Future Internet doi: 10.3390/fi18110627

Authors: Alessandro Bocci Giuseppe Bisicchia Antonio Brogi

Quantum algorithms often require many repeated circuit executions to estimate observables, guide optimization, or verify results. These repetitions, known as shots, consume quantum hardware resources and can make otherwise promising computations expensive to run. A growing body of research aims to reduce this cost, but comparing the proposed techniques is difficult: methods often optimize different quantities, rely on different assumptions, and report improvements under different conditions. This scoping review examines how shot requirements can be reduced or controlled across gate-based quantum computing. We organize existing approaches according to three fundamental decisions: how to design circuits and measurements, how to allocate shots, and how to determine when enough measurements have been collected. We examine the statistical guarantees of these approaches and the additional resources they may require, including circuit depth, coherent queries, and classical reconstruction costs. Although the reviewed methods report reductions in shot requirements or improvements under fixed shot budgets, most have been evaluated through theoretical analysis or simulation, leaving their practical performance on physical quantum hardware less well understood. We identify gaps in current evaluations and discuss how more consistent comparisons and hardware validation could help assess the benefits of shot optimization in practice.

Future Internet doi: 10.3390/fi18110626

Authors: Byron David Aguilar-Tigre Nancy Magaly Loja-Mora Jennifer Célleri-Pacheco

Support request management in public institutions is often fragmented, limiting oversight and management indicators. This study developed a natural language processing module for the automated classification of technical and information technology support requests, integrated into the institutional email system of a provincial government in Ecuador. Using a Design Science Research approach, the module was evaluated against a keyword-based baseline and a human consensus reference across three analyses: a primary corpus of institutional emails, a complementary validation of in-person requests, and an aggregate analysis of both channels. The module achieved significantly higher performance than the baseline in the email corpus (macro-F1 0.733 vs. 0.412; p = 0.0001) and in the aggregate analysis (macro-F1 0.713 vs. 0.510; p = 0.0002), while for in-person requests, no statistically significant difference was detected between the two methods (macro-F1 0.533 vs. 0.635; Δ = −0.101; 95% CI [−0.211, +0.120]; p = 0.221), with the keyword baseline nominally higher. Inter-rater agreement was substantial to almost perfect across subsets. These findings indicate that the module outperformed the implemented keyword baseline on institutional email. A supervised classifier and a second language model showed no detectable difference from the module on anonymized text.

Future Internet doi: 10.3390/fi18110625

Authors: Abdullah Khan Ayesha Siddiqa Wazir Zada Khan Qazi Emad Ul Haq

The lack of authentication and message integrity mechanisms in the Controller Area Network (CAN) exposes in-vehicle communication to flooding, fuzzing, spoofing, and replay attacks. This study presents a compact dual-stream causal depthwise-separable Temporal Convolutional Network (TCN) for five-class CAN intrusion detection. A global stream processes 32-frame windows with 11 CAN-derived features, while a same-ID stream encodes 16 occurrences using 28 causal history features. Residual temporal encoders, average and last-step pooling, learned gated fusion, and hierarchical Normal/Attack and attack-type classification are combined in a 19,685-parameter model. The model is developed separately on the Driving and Stationary preliminary submission recordings of the Car Hacking: Attack and Defense Challenge 2020 dataset using chronological 60% training and 20% validation partitions, with the final 20% reserved from model testing. Across five training seeds, validation macro-F1 reaches 98.40 ± 0.08% and 97.50 ± 0.37% for Driving and Stationary, respectively. Corresponding Spoofing F1-scores are 99.63% and 99.70%; Replay remains more challenging compared to other attacks, achieving mean F1-scores of 93.30 ± 0.42% and 88.40 ± 1.74%, for Driving and Stationary, respectively. Analysis of the validation confusion matrices shows remaining confusion between Replay and Normal traffic. In zero-shot evaluation on AutoHack_2025 dataset, mean macro F1-scores decreased to 40.20% and 34.97% for the Driving-trained and Stationary-trained models, respectively, with substantial degradation in Spoofing and Replay detection. These results demonstrate strong five-class detection under chronological within-recording validation while identifying cross-dataset robustness as the principal limitation of the current dual-stream temporal representation.

Future Internet doi: 10.3390/fi18110624

Authors: Kakim Sagindykov Nazym Mukusheva Dinara Tokseit Altyn Urynbassarova Shyryn Tutkyshbayeva

Decentralized cyber–physical systems (CPSs) improve scalability and fault tolerance but expose local controllers and neighbor-to-neighbor coordination to integrity, availability, and multi-agent attacks. This study presents a PRISMA-guided systematic mapping of cyber-resilience methods for decentralized CPSs and integrates the evidence into a taxonomy and assessment framework. A Scopus search exported on 14 September 2026 yielded 366 records; 337 primary records were screened, 163 underwent eligibility assessment, 79 were included, and 72 met a six-criterion quality threshold for Core synthesis. Power systems and microgrids represented 55.6% of Core studies, false-data/deception attacks were the largest primary attack family (31.9%), and only 18.1% evaluated explicit multi-vector or hybrid attacks. Simulation-only validation dominated (70.8%), while 29.2% provided strong experimental, real-time, hardware-in-the-loop, laboratory, or co-simulation evidence. The synthesis shows comparatively mature work on resilient control, detection, observer-based compensation, and consensus but limited evidence on recovery, self-healing, safety preservation, cross-domain transfer, and reproducible multi-vector evaluation. The proposed framework connects attack models, topology assumptions, resilience mechanisms, validation strength, and measurable outcomes, providing a traceable basis for comparing methods and designing stronger future evaluations.

Future Internet doi: 10.3390/fi18110623

Authors: Leandro Pazmiño-Ortiz Fernando Vinicio Becerra Camacho

Single observations can mischaracterize dynamic web tracking and browser fingerprinting. We measured privacy- and security-relevant indicators across a fixed, purposively curated frame of 947 Ecuador-associated organizational domains using Blacklight on 25 June, 25 July, and 25 August 2026. A completion-aware Playwright pipeline preserved the original JSON artifacts, acquisition provenance, SHA-256 hashes, quality-control results, and contextual taxonomy. Successful JSON export rates were 58.0%, 55.8%, and 59.0%; 500 domains succeeded in all three rounds, 497 satisfied measurement-level quality control in all three rounds, and the strict repeated-measures panel contained 381 domains (40.2% of the target frame) after the contextual-classification-confidence criterion was applied. Within the strict panel, six primary binary indicators varied by less than one percentage point, and none changed significantly after Holm correction (all adjusted p≥0.812); three-round exact agreement was 96.6–99.5%. The 497-domain sensitivity analysis produced closely similar results. Attrition was sector-dependent, so the strict panel is a complete-case analytical population rather than a compositionally representative subset of the original frame. A prospective direct-browser runtime-corroboration campaign on 120 stratified domains yielded 118 Firefox artifacts and 109 quality-valid observations, independently recording third-party domains, cookies, CSP headers, tracker-list matches, and fingerprinting-relevant API activity. Because contemporaneous Blacklight artifacts were not collected, this branch provides runtime corroboration rather than paired accuracy validation. The findings support traceable short-window audit prioritization and repeated-measures inference, but they do not establish national prevalence, long-term website stability, or domain-specific LOPDP violations.

Future Internet doi: 10.3390/fi18110622

Authors: Asel Abdullayeva Gulzada Mussapirova Dinara Kassymova Gabit Bakyt Kuralay Yussupova Dariga Kushtayeva

Freight logistics on the Trans-Caspian International Transport Route (Middle Corridor) suffer from long dwell times, limited cargo visibility and frequent incidents. This study quantifies how a digital escort of freight, built on a four-layer 5G-IoT-AI architecture, could change the key performance indicators of the Kazakhstan section of the corridor. Each architectural layer was mapped to a measurable mechanism and coefficient, and three segments of a Kazakhstan freight forwarder’s chain were simulated, including the railway corridor (Monte Carlo simulation, N = 10,000), the port of Aktau (discrete-event simulation) and China–Europe transit, compared with a route attractiveness index. In the baseline digital scenario, railway lead time fell by 6.3% (Cohen’s d = 1.26) and expected incidents per container by 35.0%. At the port, operational errors decreased by 34.8%, whereas dwell time decreased by only 3.2% because ferry waiting time remained the dominant contributor to overall port dwell time.r. A factorial analysis showed that AI-supported processing drives lead time, IoT monitoring drives incidents and variability, and 5G affects only response time. With the current monitored share of about 35%, the lead-time gain fell below the practical-relevance threshold. The Kazakhstan route ranked first for 95.8% of 100,000 random weight vectors and remained first when the visibility criterion was removed.

Future Internet doi: 10.3390/fi18110621

Authors: Alessandro Raschella Michael Mackay

Fifth-generation (5G) wireless networks have been key enablers for information societies over the last few years [...]

Future Internet doi: 10.3390/fi18110620

Authors: Basker Palaniswamy Paolo Palmieri

Online platforms and IoT back-ends face abrupt traffic surges. A tightly coupled monolith keeps response times short but becomes fragile under peak load, whereas loosely coupled microservices scale and isolate faults at the price of extra network communication and infrastructure. We present ATLAS (Adaptive Traffic-aware Loose–tight Architecture System), which changes at runtime which modules execute in-process and which run as isolated services. An LSTM forecasts traffic, an Isolation Forest detects anomalies in aggregate telemetry, and a binary optimisation model with a switching cost, solved exactly for small module sets, selects tight, hybrid or loose coupling (all reported results use this exact solution; a linearised heuristic for larger module sets is assessed offline only); an explicit switching protocol hands module state over between hosts and a seven-step self-healing protocol recovers failed modules. We evaluate ATLAS with a released simulator driven by synthetic workloads and with a process-level prototype. On held-out traffic, the trained LSTM reaches a five-minute MAPE of 14.8%. In 30 paired simulated runs with a flash sale, a resource anomaly and a service failure, ATLAS achieved a p99 latency of 90 ms against at least 128 ms for four autoscaled microservice baselines, 51% fewer SLA violations, no failed requests and 20% lower cost (p<0.01, Wilcoxon), with a similar recovery time. This advantage holds under the simulated conditions only: ATLAS performed worse than the baselines with pod start-up above about a minute, ±30% service-demand deviations, repeated failures, and in several combinations of adverse conditions, and its anomaly detector improved results only when an anomaly coincided with high load. The single-host prototype hands a stateful module over with a 1.4 ms pause and recovers from a process crash in 325 ms (process level, without container restart, service mesh or inter-node communication) without failed or lost requests. In a simulated connected-vehicle case study, ATLAS achieved the lowest latency (10.2 ms at peak against 20.8 ms for static microservices) and the highest SLA compliance (99.79%), at a slightly higher cost than static microservices. Smart-hospital and smart-factory scenarios illustrate further applications. Validation on production Kubernetes clusters with real traces remains future work.

Future Internet doi: 10.3390/fi18110619

Authors: Geovanny Cudco Franklin Flores-Ayala José Belisario Vera-Vera Carlos Alberto Velez-Curillo Renzo Arturo Portilla-Arias

Road traffic crashes remain a major public safety concern in Ecuador, particularly on the General Rumiñahui Highway, which is a high-demand mountainous corridor characterized by a complex road geometry, variable traffic conditions, changing environmental conditions, and limited proactive early warning capabilities. This study proposes a hybrid Internet of Vehicles (IoV)–multi-access edge computing (MEC)–artificial intelligence (AI) framework and assesses its technical and contextual feasibility through an ex ante analysis of the target corridor. The three-layer architecture integrates roadside, vehicular, and environmental sensing for data acquisition; MEC-enabled local processing, YOLOv8-based visual perception, and V2V, V2I, and V2P communications at the edge; and cloud-based spatiotemporal crash risk estimation using a CNN-LSTM model. ISO 37106:2021 and ISO 37122:2019 are adopted as design references for interoperability, smart city integration, and mobility-related monitoring. The feasibility assessment combines evidence from functionally comparable IoV, computer vision, predictive modeling, and MEC/V2X studies with an analysis of candidate Ecuadorian institutional data sources and corridor-specific deployment constraints. Literature-derived performance metrics are used only as feasibility evidence and are not interpreted as empirical measurements or the expected performance of the proposed framework. The results indicate that the architecture is technically plausible and contextually feasible under explicit conditions, providing a structured basis for Phase II digital twin evaluation followed by an instrumented field pilot under real operating conditions.

Future Internet doi: 10.3390/fi18110618

Authors: Abdelrahman Radwan Salih Alwassiti Sama Armouti Mohammad K. S. Ma’aitah Mohammad Hamdan Kassymova Dinara Ala’a Alshubbak Adnan Manasreh

Language is fundamental to social interaction and communication, and assistive technologies can help reduce barriers between Deaf and hard-of-hearing individuals and hearing communities. This study presents a standalone edge-AI device based on Raspberry Pi 5 for detecting selected Arabic Sign Language (ArSL) gestures and displaying corresponding written text. The system integrates an autofocus camera, YOLO-based gesture detection, text buffering, and a 128 × 64 OLED display capable of presenting Arabic text. YOLOv5m, YOLOv8m, and YOLOv9m were compared as candidate detectors, and a custom dataset of 1881 images was developed for nine gesture classes and background examples. The dataset was partitioned by signer, with three signers used for training, one for validation, and one held out for testing. Across three YOLOv8m training runs, the mean mAP@0.5 was 0.9583. On Raspberry Pi 5, the completed pipeline achieved a measured throughput of 3.52 frames per second (FPS). Live-camera experiments assessed performance across distance, illumination, background, viewpoint, hand orientation, and signing speed, identifying conditions that most affected detection. The results demonstrate the feasibility of a self-contained, interactive edge system for detecting predefined ArSL gestures and displaying buffered text, providing a practical foundation for expanding the vocabulary and developing continuous-sign recognition.

Future Internet doi: 10.3390/fi18110617

Authors: Thomas Loruenser Stephan Krenn

Progress in information and communication technology, together with the demand for digital transformation, continues to produce networks, systems, and platforms that enable collaboration and data-driven workflows at scale and beyond established trust boundaries [...]

Future Internet doi: 10.3390/fi18110616

Authors: Khaled Hmidi Sameh Najeh Ammar Bouallegue

UAV, mounted aerial base stations offer remarkable deployment flexibility for 6G Non Terrestrial Networks (NTNs), yet their hovering energy demands present a critical sustainability bottleneck, particularly as network topologies continuously evolve in practice. As a first step, we develop MARL-ESA (Energy-Saving Adaptation), a Multi Agent Reinforcement Learning (MARL) framework achieving 48% energy savings in static UAV assisted heterogeneous networks (HetNets) compared with always-on operation through centralized training with decentralized execution (CTDE). However, MARL-ESA remains inherently topology specific, requiring over 2000 retraining episodes whenever UAV positions or traffic patterns change. To overcome this limitation, we propose Meta-MARL-ESA, a meta reinforcement learning extension that enables zero shot generalization across unseen network configurations. The problem is formulated as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) over a task distribution τ∼p(T), with a permutation invariant Graph Neural Network (GNN) policy trained via Model Agnostic Meta Learning (MAML). Simulation results demonstrate that Meta-MARL-ESA achieves 12× faster convergence on new topologies, preserves 82% of optimal performance in zero shot deployment, improves energy efficiency (EE) by 21.5% over converged baselines in out of distribution (OOD) scenarios, and recovers to 98% of optimal performance within 8 decision epochs during dynamic UAV repositioning events.

Future Internet doi: 10.3390/fi18110615

Authors: Syed Ijlal Haider Mohammad Nouman Raja Ujjan

Model Context Protocol (MCP) clients may approve a Tool whose authenticated, client-visible contract later changes. Directional Contract Transition Revalidation (DCTR) retains grant-bound operational and last-human-reviewed baselines, canonicalizes contracts with RFC 8785/JCS, and assigns L0–L4 renewal actions to typed directional changes. Because a maximum of individual path levels can miss jointly introduced behavior, the final proposed DCTR-IA-2.0 decision rule requires a hash-bound whole-contract assessment for multi-path or coupled schema changes in either the local or cumulative delta; unresolved cases remain pending human reapproval. With official MCP Python SDK clients, this gate withheld the original compound counterexample in simultaneous and sequential forms and a four-case SDK outcome matrix (one reusing the paired field pattern and three testing different mutation types); three benign controls also incurred review. On 20 natural N2 Tool transitions (43 changed paths, 14 maintainer clusters), a reproducible structural replay routed 14 contracts to whole assessment, without establishing natural interaction severity or unnecessary review rates. Independent expert agreement was 28/34 paths in the frozen held-out study (weighted κ = 0.720), 18/21 in the separate pilot, and 37/48 (κ = 0.710 on 47 ordinal pairs) across 20 selected N3 transitions in eight repositories; jointly assigned L3 paths occurred in 11 clusters across five repositories, with ten subsequently documented casewise panel agreements and one separately documented DEFERRED case; the author-submitted post-review tally classifies 47/48 (L0 = 3, L1 = 5, L2 = 19, L3 = 20, L4 = 0) without assigning a level to N3K0023 or claiming independently authenticated expert sign-offs or natural L4. A separate selected Study B yielded 10/10 agreement distinguishing six adapted R4/L4 poisoning stimuli from four disclosed R3/L3 controls. These controlled assessments and enforcement tests do not establish autonomous semantic risk detection, reduced production reapproval burden, or prevented harm. ETDI and Microsoft contrasts are comparisons of policy design only.

Future Internet doi: 10.3390/fi18110614

Authors: Sarmad M. Hadi Walled Khalid Abdulwahab Ali A. Muhamed Ali Salim Abdulateef Mohammed Ali Mohammed Imad Aal-Nouman

Application-level evidence is needed to evaluate communication-related delay, processing, reliability, and hosting cost when selecting cloud virtual machines for geographically distributed database-backed applications. This study presents a communication-aware Java TCP/MySQL benchmarking and decision-support framework in which the Java server and MySQL database are co-located on each server VM. The original sequential request campaign compared AWS t2.medium, Azure Standard_D2s_v3, and GCP e2-medium across three client-hosting providers and three client locations: Tokyo, London, and São Paulo (129,600 completed requests). The evaluated metrics included total response time (TRT), server/database processing, a derived application-layer communication-related residual, and request completion, alongside three Communication-aware Cloud Hosting Score (CCHS) scenarios. Under the tested conditions, Azure recorded the lowest overall mean TRT (404.67 ms), while AWS achieved the highest mean CCHS under the balanced (S2) and cost-aware (S3) scenarios. A separate supplementary campaign evaluated three independently provisioned server VMs per provider, AWS-hosted clients, and concurrency levels of 1, 10, and 30 (121,500 successful requests). The mean TRT values were 226.13 ms for AWS, 236.68 ms for Azure, and 228.19 ms for GCP. Because deployment configurations and workloads differed, the campaigns were analyzed separately. The results characterize the selected deployments rather than establishing provider-wide rankings or directly measuring network-path latency.

Future Internet doi: 10.3390/fi18110612

Authors: Jyothsna Laxmi Saripalli Hari Hara Babu Saripalli Leonel Lagos Pranav Gangwani Jayesh Soni Alexander Perez-Pons Himanshu Upadhyay

Acoustic sensor nodes protecting logistics sites use fixed-label classifiers, so adding a class requires new data, retraining, and redeployment of every node. Audio–language models promise to add classes by text description. We evaluated three frozen contrastive audio–language models on 25,599 events (15,241 gunshots) and on an external field set, and measured their cost on six edge platforms. On supplied clips, a vocabulary-independent score calibrated on separate data for a 5% false-positive target recalled 79% to 84% of gunshots at realized rates of 4.8% to 5.6% against unseen background sounds. On field recordings from an unseen domain, separation held on full clips (area under the ROC curve 0.95 to 0.97), but the clip-level false-positive rate rose to 7% to 11%, and only 24% to 25% of gunshots were detected end to end because the onset gate missed distant shots. Fine-grained zero-shot classification was weak: balanced accuracy was 0.53 against 0.50 for supersonic versus subsonic fire and 0.05 against 0.04 for firearm identity. A linear probe reached 0.82 at the regime level, but 0.59 with calibers held out. A firearm class added by one sentence reached 6.4% accuracy, against 54% with twenty labeled examples.

Future Internet doi: 10.3390/fi18110613

Authors: Oghenetega C. Owivri Kennedy O. Okokpujie Samuel A. Daramola

Private keys are the single point of failure in blockchain systems; if lost, assets are irrecoverable, and if stored, they are exposed. This paper presents a template-free biometric cryptosystem that regenerates a 256-bit private key directly from a fingerprint, storing neither the biometric template nor the key. A Siamese convolutional network trained with a two-stage composite loss produces a 1024-bit binary representation; an adaptive, margin-based binarisation with stability-aware selection retains 288 reliable bits; and a block-wise BCH fuzzy extractor regenerates the exact key, while a layered scheme of Scrypt, AES-256-GCM, and Shamir secret sharing protects the helper data. Evaluated on the PolyU fingerprint dataset under a strict cross-session protocol, the system reduces the mean intra-user bit error rate from 8.13% to 2.35%, attains an equal-error rate of 6.27% and an area under the curve of 0.9831, and achieves a 95.83% key-reconstruction success rate with a 0% false key-acceptance rate. The 288-bit key has an estimated effective min-entropy of 269 bits, and helper-data leakage is negligible (mutual information 0.0132). The results show that template-free, cross-session key generation is feasible at the 256-bit level required for blockchain.

Future Internet doi: 10.3390/fi18110611

Authors: Amir El-Azazy Mahmoud M. Badr Raafat El-Kammar Hala Abd-Elkader Ahmed Fawzy

Offline Egyptian Arabic automatic speech recognition (ASR) requires balancing recognition accuracy, model size, and edge device performance. This study investigates structured feed-forward network (FFN) compression and recovery fine-tuning of a 114.6M-parameter FastConformer Hybrid RNN-T/CTC model, followed by offline Raspberry Pi 4 deployment. Reducing FFN width from 2048 to 1792 produced a 105.7M-parameter model, a 7.78% reduction. Structured compression initially degraded recognition accuracy, while subsequent recovery fine-tuning restored performance without restoring the removed FFN channels. On a frozen 10,361-utterance test set, the compressed-and-recovered model achieved 25.01% word error rate (WER) and 11.96% character error rate (CER), compared with 25.88% WER and 12.58% CER for the reference model. A dense control given the same recovery training reached 24.81% WER and 11.87% CER; its paired WER difference from the compressed-and-recovered model was −0.20 percentage points (95% confidence interval −0.44 to +0.04; one training run per model). The earlier advantage over the reference therefore does not indicate a compression-specific benefit. On Raspberry Pi 4, the reference and compressed-and-recovered models achieved aggregate real-time factors of 0.470 and 0.466 and mean latencies of 1.185 and 1.176 s, respectively. The parameter reduction yielded no measured Raspberry Pi speed or memory advantage; traffic-command recognition was not evaluated.

Future Internet doi: 10.3390/fi18110610

Authors: Luis-Aaron Zurmühlen Peter Weber

Web search is a primary gateway for accessing digital information, and artificial intelligence (AI) fundamentally reshapes both search infrastructure and search engine optimization (SEO) practices through which organizations achieve online visibility. Despite this dual transformation, no study has systematically mapped the position of SEO research within this field. This study addresses this gap through keyword co-occurrence analysis of 2273 publications from Scopus, Web of Science, and IEEE Xplore (2015–2025). The analysis identifies five thematic clusters that overlap in temporal regards and represent the technical aspects by spanning retrieval, semantic processing, knowledge extraction, and large language model-based architectures, concluding in conversational search. However, SEO-specific practice-oriented terminology remains sparse and confined to a single cluster. Building on these findings, the study develops a research agenda comprising seven research questions across three directions: evaluating AI-adapted SEO strategies and optimization for generative search, conceptualizing visibility under AI-mediated search, and examining stakeholder adaptation and value distribution within the search ecosystem. These findings provide a foundation for future research addressing the growing gap between the field’s technical and SEO-specific dimensions.

Future Internet doi: 10.3390/fi18110609

Authors: Mohamed Ahmed El Ghazouly Duaa Zuhair Al-Hamid Omar Gadelmawla Hamid GholamHosseini Mirza Mansoor Baig Shereen Afifi

Real-time monitoring of muscle fatigue could support safer strength training, yet most wearable systems depend on cloud processing, which adds latency and exposes physiological data. This work presents and evaluates an edge-based prototype for biometric monitoring during biceps curls. A wearable node integrating a nine-axis IMU (MPU-9250) and a photoplethysmography heart-rate sensor (MAX30102) streams data from an ESP32 microcontroller over UDP to a Raspberry Pi 5, which performs feature scaling and inference with a compact feedforward neural network (two hidden layers of 64 and 32 units) trained on a publicly available biceps-curl dataset. The classifier takes a 10-element feature vector of IMU, heart-rate and acceleration-derived inputs and outputs three fatigue levels. Under leave-one-subject-out cross-validation on ten participants, the three-class classifier attained a mean accuracy of 0.77 (SD 0.25; 95% CI 0.59–0.95), with a strongly bimodal per-subject distribution (0.40–0.98). Deployed on the edge node and evaluated end-to-end over 5172 annotated windows under a binary no fatigue versus fatigue formulation, the model recovered 91% of non-fatigued but only 10% of fatigued windows (balanced accuracy 0.51, Matthews correlation 0.02), which is no better than chance and below a majority-class baseline. In its current form the classifier is therefore not a reliable fatigue detector; we analyse candidate causes, including training-side class skew and differences between the source-dataset and deployment sensors, and compare them against majority-class and chance reference points. On the systems side, the pipeline runs fully locally with 10 ms model inference per sample, 42 ms gateway processing and a mean wireless round-trip time of 180 ms after rate-limiting the link. The implemented physical, communication and edge-inference layers are designed as the sensing foundation for a future Digital Twin, which is not implemented in this work.

Future Internet doi: 10.3390/fi18110608

Authors: Alexandros Kodonas Stavros Kalogiannidis Athanasios Papavasileiou Olympia Papaevangelou

This cross-sectional survey examines the associations of smart-device usage, personalization, and IoT integration with perceived customer engagement among 392 Greek business and digital-marketing professionals. The leading reported functions were instant communication (35.2%), tailored recommendations (37.5%), and real-time data collection (30.6%); privacy concerns were the most frequently reported adoption barrier (34.2%). The three-predictor regression was statistically significant: F(3, 388) = 138.713, p < 0.001, with R2 = 0.517 and adjusted R2 = 0.514. The reported standardized coefficients were positive for smart-device usage (β = 0.381), personalization (β = 0.304), and IoT integration (β = 0.265). The estimates concern professional perceptions and remain provisional because sampling and measurement diagnostics are incompletely documented; they do not establish causal effects or actual consumer behavior. The implications concern relevant customer interactions, interoperable infrastructure, transparent data practices, and privacy-conscious personalization. Managers should evaluate connected-marketing investments alongside implementation cost and infrastructure readiness. Further research should document sampling and measurement procedures, reproduce the model with respondent-level data, and evaluate generalizability across countries and industries.

Future Internet doi: 10.3390/fi18110607

Authors: Ajay Bandi Benitha Sri Panchagiri Rushika Reddy Pingili Akash Thanneeru Jonathan Grant Geren Popin Bose Roy

RAISE (Real-time AI Insights for Sentiment and Engagement) is a preliminary framework that combines speech transcription with Whisper, speaker diarization with pyannote.audio, text-based emotion classification with DistilRoBERTa, and feedback generation with GPT-3.5. We examined the workflow using 53 audio clips: 12 author-recorded conversations, 23 YouTube clips, and 18 clips from public emotion-labeled datasets, including RAVDESS and IEMOCAP. The experiments provide an initial demonstration of how these components can be integrated and identify speaker diarization as a major source of processing latency. Because the dataset is small and the evaluation is exploratory, the results do not establish generalizable accuracy or reliable real-time performance. Further evaluation using larger, independently labeled datasets and longer conversations is needed.

Future Internet doi: 10.3390/fi18110606

Authors: Mohammad Nouman Syed Ijlal Haider Raja Ujjan

Tool poisoning can induce Model Context Protocol (MCP) agents to perform harmful host-side actions through manipulated tool metadata or invocation arguments. This paper studies a client-side least-privilege capability-inference mechanism that translates concrete tool names and recursively structured arguments into six behavioural categories before policy lookup. On a review-driven audit of 27 calls labelled from documented official MCP reference-server semantics, the corrected mapper obtains 27/27 exact category sets. On 16 author-instantiated, external-provenance MCPTox cases, it recovers all specified minimum capabilities with three additional conservative categories; neither set constitutes blind generalisation evidence. A separate policy matrix yields 27/27 expected decisions. An independently implemented declared-capability policy comparator agrees with B4 on 170/172 decisions across four profiles, including 43/43 under limited-network (0% ASR; 59.26% benign-call allow rate). Hardened launch validation rejects eight malformed or bypass specifications. Twelve live stdio checks verify tool discovery, two benign workflow outputs, denied-write non-execution, error propagation, ASK denial, symlink protection, and upstream-crash failure. The 20-workflow authorisation replay is reported separately from executed task success. The contribution is automatic name/argument-to-capability inference and bounded enforcement evidence, not universal interception or confinement of malicious server internals.

Future Internet doi: 10.3390/fi18110605

Authors: Mohamed Benaida Fayçal Hamdi Mahmud Mansour Ahmad B. Alkhodre Abdullah Alshanqiti Tanweer Alam

As governments expand artificial intelligence (AI) in public services, user trust may depend on the governance safeguards surrounding its deployment. This exploratory study combines a documentary analysis of 16 AI governance dimensions in the United Kingdom and Saudi Arabia with a survey of 100 GOV.UK and Absher users recruited through non-probability sampling (50 per group). The documentary comparison identifies contrasting governance arrangements, with a more distributed, sector-led approach in the United Kingdom and more centralised coordination in Saudi Arabia. Using an adapted 12-item trust questionnaire, surveyed Absher users reported higher average scores than surveyed GOV.UK users (3.47 versus 3.29), with the largest item-level differences concerning suspiciousness, wariness, and perceived integrity. Given limited internal consistency and unestablished English–Arabic measurement equivalence, these differences are interpreted as exploratory item-score patterns. Respondents’ awareness of national AI governance arrangements was not assessed, so structural assurance is discussed as a theoretical lens rather than a tested explanation. The study contributes a structured governance comparison, descriptive evidence of trust-item patterns in two selected platform-user groups, and a five-component framework addressing transparency, accessibility, accountability, enforcement visibility, and continuous trust monitoring.

Future Internet doi: 10.3390/fi18110604

Authors: Ayoub Alsarhan Mamoon Obiedat Malek Barhoush Mohammad Mansour Al-Khasawneh Sami Aziz Alshammari Nayef H. Alshammari Khalid Hamad Alnafisah Mohammed Aleinzi

Road networks are routinely exposed to partial capacity losses caused by incidents, work zones, flooding and pavement failures, and planners need fast, trustworthy estimates of how such losses propagate through route choice. This paper develops a framework for user-equilibrium (UE) traffic networks with degradable link capacities that combines variational analysis with a learned surrogate. Writing the Beckmann potential in the coordinates ua=ca−p makes it affine in u, so standard results (the concavity of a pointwise minimum of affine functions and Danskin’s theorem) show that the optimal potential is concave, non-decreasing and differentiable, with a gradient available from one equilibrium solution. The resulting criticality index equals p/(p+1) times the congestion delay of each link; the theory adds to this classical indicator a derivative interpretation and certified bounds. We derive two-sided certificates on the potential change for any degradation, the identity TSTT=(p+1)V−pFFT, a solver-free certificate for the total system travel time (TSTT), and an additive certificate for restoring sets of links, whose per-link weight is the damaged-state delay of the link times a severity factor, and we use these quantities as features of an envelope-informed surrogate (EIGS), a gradient-boosted model on per-link envelope features diffused over the line graph. Experiments on the Sioux Falls and Anaheim benchmarks show where the results are useful and where they are not. The potential certificate is within a factor of 1.45 of the exact change at 25% capacity loss on Sioux Falls (1.03 on the 100 most critical Anaheim links) but loosens quickly with severity (factor 3.28 at 50% loss on Sioux Falls; 574 and 74 at 90% loss), and the solver-free TSTT bound exceeds the true change by a median factor of 3.8–12 in the multi-link Monte Carlo scenario sets. No single cheap indicator ranks link criticality best on both networks. On Sioux Falls, EIGS predicts the TSTT increase with R2=0.942 (0.913–0.960 over ten splits) against 0.919 for ridge regression and 0.924 for gradient boosting on the diffused severities, and with R2=0.909 against 0.862 on a spatially correlated test set generated after the model was fixed. Under shifts in demand (±15%) and nominal capacity, refreshing its envelope features from one equilibrium of the shifted network gives R2=0.86–0.91; using the same equilibrium to rescale the predictions of raw-input baselines gives them 0.84–0.89, so the gain from the per-link envelope features is modest. On Anaheim, whose losses are extremely heavy-tailed, EIGS ranks scenarios well (Spearman ρ=0.92) but does not predict their magnitude reliably, while ridge regression with a cross-validated penalty ranks them poorly (ρ=0.16). For restoring three of about fifteen damaged links, the certified envelope rule, which needs a single equilibrium, recovers on average 93.6% of the loss recovered by the exact optimum over all three-link sets.

Future Internet doi: 10.3390/fi18110603

Authors: Aliya Abdiraman Laura Aldasheva Akzhibek Amirova Didar Yedilkhan Moldir Azim De Mi

Quantitative assessment of information security in 5G telecommunications infrastructure is difficult because individual security metrics are heterogeneous and rarely interpreted together. Conventional multi-criteria decision-making (MCDM) applications rank alternatives by compensatory aggregation, so a violated metric can be offset by others. This study proposes a standard-traceable method in which MCDM only weights the metrics, while measured values are evaluated against tolerable values by a non-compensatory exceedance indicator and a three-level status rule. Twenty-four metrics were derived from ISO/IEC 27011:2024 controls through the ISO/IEC 27004 measurement chain and grouped into three domains. The analytic hierarchy process (AHP) weighted three quality criteria of a metric (significance 62.6%, correctness 28.9%, measurability 8.5%), and the weighted aggregated sum product assessment (WASPAS) method used them to derive the metric weights, cross-checked by the fuzzy technique for order of preference by similarity to ideal solution (Fuzzy TOPSIS). Validation combined synthetic deterioration scenarios, Monte Carlo analysis and comparison with alternative MCDM methods. In three 12-month scenarios, the exceedance indicator flagged each violation in the month it occurred, whereas compensatory aggregation lagged by three to six months or did not react. The domain status remained stable in 85.3–100% of Monte Carlo draws under perturbed tolerable values and measurements. Compared with conventional MCDM-based aggregation, the advantage lies in traceable metric selection, normalization against tolerable values and the non-compensatory status rule.

Future Internet doi: 10.3390/fi18110601

Authors: Berenice Arguero Jessica Shagñay Christian Tipantuña Germán V. Arévalo Pedro J. Reyes-Iglesias

This paper presents a Systematic Mapping Study of Machine Learning techniques applied to Optical Access Networks, with an emphasis on Passive Optical Networks. From 605 Scopus-indexed publications, 112 primary studies were selected and analyzed according to learning model, application domain, dataset characteristics, validation maturity, reported limitations, and future research directions. The results show that research is strongly concentrated on resource management, particularly Dynamic Bandwidth Allocation, traffic prediction, and performance optimization, whereas security, resilience, and operational robustness remain comparatively less represented. A substantial reproducibility and maturity gap is also identified: only four primary studies (3.57%) use publicly accessible datasets, 94.64% of the corpus remains at the simulation-based or experimental level, and no operational deployment was identified. To complement the systematic mapping, we present an illustrative proof of concept that evaluates ML-based traffic classification as a potential support component for traffic-aware DBA. The originality of this mapping lies in jointly examining ML application domains, dataset accessibility, reproducibility, and technological maturity within the specific context of optical access networks. The findings provide a structured basis for identifying deployment barriers and prioritizing representative datasets, standardized benchmarks, and operational validation in future PON research. Using a controlled synthetic dataset, three supervised classifiers achieved weighted F1-scores close to 0.80, with Random Forest exhibiting the lowest measured inference latency. These results provide descriptive evidence under the evaluated conditions and do not constitute validation of an end-to-end DBA mechanism or an operational PON deployment. Overall, the findings highlight the importance of public and representative datasets, standardized benchmarks, validation under realistic network conditions, lightweight and explainable models, and integration with programmable SDN/NFV-enabled PON control architectures for advancing ML applications toward practical deployment in optical access networks.

Future Internet doi: 10.3390/fi18110602

Authors: Yuta Goto Joe Hasei Tatsuo Nakajima

Some patients require long-term hospitalization, which restricts their opportunities for social connection and participation in important events or daily lives. To mitigate this issue, we developed an asymmetric 360-degree telepresence system requiring minimal hardware. At the remote site, the user wears a virtual reality head-mounted display to be immersed in the environment. At the local site, a 360-degree camera captures the surroundings, while a humanoid avatar displayed on a rotatable smartphone enables local users to perceive the remote user’s gaze and gestures. In an exploratory, controlled laboratory evaluation (N=16, 8 pairs), remote users reported positive preliminary spatial presence while social presence and non-verbal task completion times showed descriptive variations across visual feedback conditions without statistically significant differences, rather than demonstrating task efficiency gains or statistical equivalence to conventional video calls. Furthermore, in an exploratory single-case fieldwork (N=1), a hospitalized adolescent used our prototype system to remotely view her graduation ceremony. In a post-event subjective questionnaire, the participant reported favorable personal impressions regarding the opportunity and experience to view the ceremony online, although severe cellular network congestion limited smooth video streaming to approximately 20% of the session, and standardized clinical or psychological outcomes were not measured. Overall, this work highlights the potential of lightweight asymmetric telepresence as an initial prototype, while identifying key technical, network, and methodological challenges that must be addressed for future real-world deployment.

Future Internet doi: 10.3390/fi18110600

Authors: Alina Georgiana Badea Irina Andra Tache Lidia Băjenaru Severus Constantin Olteanu Silviu Stanciu

Accurate delineation of cardiac structures in cardiac magnetic resonance (CMR) images is a prerequisite for quantifying cardiac function. This paper presents an end-to-end pipeline for automated cardiac segmentation and five-class pathology classification, evaluated on the 150 patients of the Automated Cardiac Diagnosis Challenge (ACDC) dataset. Three U-Net-based networks were trained with a hybrid cross-entropy-Dice loss selected through an ablation study and combined through soft-voting and weighted ensembles. On the official test set, the weighted ensemble achieved a mean Dice score of 90.27% at the slice level and 86.23% in volumetric evaluation, matching the best individual network and significantly outperforming the other two. Biomarkers extracted from the ensemble masks agreed closely with ground-truth-derived values (left ventricular ejection fraction: intraclass correlation 0.975) and were used to train nine machine learning classifiers under nested cross-validation. The four best classifiers achieved macro F1 scores between 88.73 and 89.14%. A cross-source evaluation across five segmentation sources showed no significant difference between classification based on automatically derived biomarkers and on ground-truth-derived biomarkers. SHAP analysis identified class-specific biomarkers consistent with cardiac pathophysiology. The results demonstrate technical feasibility on a public benchmark; external multi-center validation is required before clinical use.

Future Internet doi: 10.3390/fi18110599

Authors: Daniel Dauda Wisdom Abayomi Olusola Agbeyangi Jose Manappattukunnel Lukose

Business intelligence (BI) systems increasingly rely on data-driven architectures to facilitate informed financial and operational decisions. However, the integration of advanced algorithms into these platforms has also enabled malicious actors to deploy sophisticated, machine-learning (ML)-orchestrated cyberattacks that easily bypass traditional security signatures. To address this vulnerability, this study proposes a hybrid defensive framework combining elliptic-curve cryptography (ECC) and ML to secure the computational and authentication infrastructure underpinning BI transaction pipelines against automated adversarial threats. A mathematical model grounded in mapping theory was developed to facilitate real-time anomaly detection and adaptive threat mitigation. Furthermore, the proposed architecture introduces a continuous user authentication protocol that evaluates login attempts against established behavioural profiles, immediately flagging divergent access patterns for secondary review. For empirical validation, the framework was evaluated using hardware performance counter (HPC) telemetry from the Canadian Institute for Cybersecurity DDoS Attack Dataset (CICEV2023) within a virtualised execution environment. Low-level execution metrics, including CPU cycles, branch instructions, and executed instructions, were analysed as an infrastructure-tier proxy to assess computational exhaustion and instruction diversion under adversarial authentication attacks. Six ML algorithms were evaluated for intrusion detection. The random forest classifier demonstrated superior performance, achieving an accuracy of 81.58%, a specificity of 89.36%, a precision of 80.00%, and an area under the curve (AUC) of 89.29%. The findings indicate that integrating ECC-based cryptographic verification with ML-driven behavioural analysis yields a resilient, multi-layered defence scheme capable of identifying and mitigating infrastructure-level cyber threats in real-time.

Future Internet doi: 10.3390/fi18110598

Authors: Vladimir Stantchev

Authorization in Internet of Things deployments assumes a reachable cloud; the cloud-fog-edge continuum breaks that assumption where decisions matter most. This article presents a delegation architecture in which authority is granted at the cloud and attenuated tier by tier down to fog nodes, gateways, and constrained devices, with caveats for topic scope, rate, geofence, and validity window. The architecture requires every tier to verify locally and offline, binds requests to the leaf key, bounds revocation by epochs and expiry, and records decisions as hash-chained evidence. A Python reference implementation realizes verification, request handling, revocation, and evidence; the device-tier verifier and the attested OP-TEE fog verifier are specified, not built. In a medical IoT scenario, CBOR-encoded chains fit CoAP without block-wise transfer up to depth 4 and a negotiated BLE notification at depth 1; on LoRaWAN only a depth-1 chain fits, at the fastest data rates without repeaters, so LoRa-class links need session-anchored authorization. Under hour-long disconnection, authorization availability is 0.93–0.98 for expiry horizons of 8 h or more, and a credential revoked while still valid stays honored for 32–33 min. On three Raspberry Pi boards, chain verification costs 327–645 μs per hop, under 4 ms at depth 6, and about 0.38 mJ per hop.

Future Internet doi: 10.3390/fi18110597

Authors: Mousa Alhajlah

The increasing availability of diverse internet of things devices has made possible an increase in zero-day attacks, but at the same time, it has brought up several issues related to data privacy, overhead communications, and computation scalability for central intrusion detection systems. Federated learning provides an effective decentralized approach for cooperative cyber threat intelligence; however, its efficiency is hindered by non-independent and identically distributed data, unstable client contribution, restricted resources of edge devices, and ineffective model aggregation. In order to overcome the aforementioned problems, this paper presents an Adaptive Trust-Aware Federated Edge Transformer (AFET) with multi-level knowledge distillation (MLKD) and hybrid Genetic Algorithm–Sequential Quadratic Programming (GA–SQP) optimization for privacy-preserving zero-day attack detection in diverse IoT environments. The AFET architecture implements an adaptive trust-aware model aggregation process that adjusts the aggregation weights based on the clients’ reliability and data quality, whereas the MLKD module implements knowledge transfer from a global teacher network to edge transformers through multi-level information sharing. In addition, the hybrid GA–SQP optimizer optimizes the parameters of both the transformer and federated learning algorithms. The experimental results performed on benchmark intrusion detection datasets prove that the proposed framework attains detection accuracy of 99.18%, precision of 98.07%, recall of 97.91%, F1-score of 98.03%, and AUC of 0.998. This is an alternative tool for privacy-preserving zero-day attack detection in heterogeneous IoT networks.

Future Internet doi: 10.3390/fi18110595

Authors: Dipanjan Adhikary Eirini Eleni Tsiropoulou

Practical deployment of uplink Non-Orthogonal Multiple Access (NOMA) suffers from synchronization errors, device-specific carrier-frequency offsets, hardware impairments, and the complexity of Successive Interference Cancellation (SIC). This paper presents an end-to-end three-user, narrowband single-carrier power-domain uplink NOMA testbed built with low-cost ADALM Pluto Software-Defined Radios (SDR) that operate without a shared reference clock. A beacon-and-response protocol is executed directly on each radio’s native ARM processor and provides low-overhead transmission coordination. The receiver addresses any residual asynchrony among the received signals through a user-specific preamble acquisition, carrier-frequency-offset estimation and correction, matched filtering, least-squares channel estimation, pilot-aided phase tracking, soft combining of repeated packet copies, convolutional decoding, waveform reconstruction, and SIC. The designed platform is evaluated over 1000 communication rounds in both single-cell and simultaneous two-cell configurations, with three transmitting users and one central receiver per cell. In the single-cell experiment, the beacon detection success rate reaches 93.17%, while the system packet error rate is only 0.45% when all users detect the beacon. In the two-cell experiment, the combined beacon detection rate is 92.55%, and the corresponding system packet error rates are 0.12% and 1.81% for the two cells. The average receiver-side processing remains below 100 ms, while end-to-end latency stays below 0.36 s despite the host communication and file-transfer overhead. The proposed implementation closes the gap between analytical NOMA theoretical research approaches and practical over-the-air deployment, and provides a reproducible foundation for scalable multicell experimentation and future integrated real-time processing.

Future Internet doi: 10.3390/fi18110596

Authors: Cosimo Melella Alejandro Arenas Tovar Manuel Díaz Pastor Federico Carvajal Aida Akbarzadeh

Security-testing platforms that rely on public clouds create tensions with data-protection requirements and digital-sovereignty objectives, yet existing cyber range (CR) architectures rarely address regulatory compliance by design. This paper proposes a reusable, compliance-aware CR method comprising a five-step process—gap analysis, principle selection, regulatory mapping, architectural instantiation, and principle-driven validation—together with five technology-agnostic design principles: sovereignty, ephemeral data minimisation, isolation at the infrastructure boundary, infrastructure-as-code (IaC)-driven reproducibility, and continuous monitoring. The method is instantiated as a vendor-agnostic dual-cluster Kubernetes architecture on on-premise OpenStack infrastructure: a persistent Red cluster hosts the offensive tooling (Kali Linux, OWASP ZAP) and a GitLab CI/CD runner, while an ephemeral Blue cluster hosts the target application together with its security-monitoring stack and is provisioned via Ansible at the start of each audit cycle and destroyed at its conclusion, so that target data does not persist beyond the audit window. A proof-of-concept deployment with a deliberately vulnerable web application (ten seeded flaws) exercises the lifecycle end to end. Isolation holds at the OpenStack Security Group layer; the Kubernetes NetworkPolicies are specified but not enforced by the CNI used in the prototype. A mapping of the framework’s principles to specific articles of the GDPR, Cyber Resilience Act, and NIS2 Directive identifies which obligation each architectural property supports. The mapping is not a determination of compliance.

Future Internet doi: 10.3390/fi18110594

Authors: Olga Torstensson Dmytro Prokopovych-Tkachenko Alona Desiatko Kostiantyn Babenko Dmytro Tyshchenko Tamara Franchuk

Artificial intelligence (AI)-native sixth-generation (6G) networks are expected to rely on agents in different administrative domains that coordinate policies, model versions and access rights through shared state. A permissioned ledger can anchor this state, but an edge agent can act on an update only when its own read node returns it. We use this verified state visibility, measured at the controlled read nodes, as both the service objective and the learning signal for block-pacing control in a permissioned Avalanche layer-one (L1) chain. The predictive components—gradient-boosted quantile regression with a split-conformal bound—are established; the contribution is the signal against which they are trained and checked. In simulated 6 h episodes with non-stationary load and transport degradations, a risk-aware controller that respects the rate limit of Avalanche Community Proposal 226 (ACP-226) reduced validator central processing unit (CPU) load by about 18% relative to static 250 ms pacing at a service-level objective (SLO) violation rate of about 2%. Trained on transaction receipts instead, the same controller violated the objective in about two of three windows. A reactive rule driven by verified reads performed comparably, so the gain stems from the learning signal rather than from learning-based control. In a case study, verified reads exposed an injected calibration fault online. All results are simulated.

Future Internet doi: 10.3390/fi18110593

Authors: Sunil Kumar Sameeksha Khare Isha Chopra Jugul Kishor Gupta Devendra Kumar Wee How Khoh Firoz Khan Vivek Kumar Sharma

The increasing adoption of Internet of Things (IoT) technologies in smart farming has improved agricultural automation and decision-making while expanding the attack surface of interconnected devices and communication networks. Intrusion detection in these environments is challenged by heterogeneous network features, severe class imbalance, and evolving attack patterns. This study proposes CCRL-LightGBM, a contextual contrastive representation learning framework that integrates TabTransformer-based feature learning, supervised contrastive refinement, and Light Gradient Boosting Machine (LightGBM) classification. The framework was evaluated on the CICIoT2023 and CIC IoT-DIAD 2024 datasets using a unified preprocessing pipeline and compared with representative machine learning and deep learning models. Under random-stratified evaluation, CCRL-LightGBM achieved an accuracy of 99.63% and F1-score of 99.81% on CICIoT2023 and an accuracy of 99.39% and F1-score of 99.68% on CIC IoT-DIAD 2024. GPU inference latency was 3.78 ms per sample under the evaluated experimental conditions. In paired comparisons with TT–LightGBM, the CCRL-LightGBM configuration achieved higher F1-scores under temporal, source-disjoint, and device-disjoint evaluation protocols, with device-disjoint F1-scores of 93.31% on CICIoT2023 and 92.56% on CIC IoT-DIAD 2024. However, performance declined under stricter partitioning protocols, highlighting the challenges of generalization to temporally shifted traffic and previously unseen source or device groups. Unadapted cross-dataset evaluation further revealed limitations in transfer performance between the two benchmark environments. These results demonstrate strong benchmark-level performance and improved results relative to the baseline under the evaluated conditions while highlighting the need for further validation using agriculture-specific traffic and deployment-level measurements before operational suitability for smart-farming environments can be established.

Future Internet doi: 10.3390/fi18110592

Authors: Bushra Alkomah Mohammed G. Almutairi Frederick T. Sheldon

Political misinformation spans fact-checked claims and news headlines, yet detection models are often evaluated within a single corpus and input representation. This study presents a metadata-flexible transformer framework evaluated on LIAR2 (19,212 retained claims) and FakeNewsNet-PolitiFact (928 headline-level news items). RoBERTa-base and DeBERTa-v3-base are evaluated using text-only and context-enhanced representations across supervised classification, cross-dataset transfer, few-shot adaptation, and robustness experiments. Across five random seeds, contextual RoBERTa achieves a pooled macro-F1 of 0.7369±0.0072, while DeBERTa with source-withheld context reaches 0.7426±0.0128. On the FakeNewsNet-PolitiFact test partition, full-context DeBERTa achieves a macro-F1 of 0.8842±0.0120, demonstrating strong supervised classification performance. Additional five-seed experiments evaluate robustness under cross-corpus overlap screening, speaker-disjoint partitions, and alternative binary label mappings. These analyses demonstrate that the benefits of contextual information extend across multiple controlled evaluation conditions. Cross-dataset transfer experiments further investigate generalization between political claims and news headlines, revealing differences in model behavior across input representations and data distributions. Five-seed forward-transfer evaluation provides additional evidence concerning the influence of contextual information on cross-corpus performance. In a complementary fixed-seed few-shot experiment, adaptation with 50 labeled target examples per class achieves 0.8480 macro-F1, representing a +0.2653 improvement over zero-shot performance. Matched few-shot experiments further examine the contribution of source-domain initialization under different target-data budgets. Overall, the results demonstrate the effectiveness of context-aware transformer models for political misinformation classification while providing a comprehensive evaluation of their robustness, transferability, and adaptation capabilities. The findings highlight the importance of contextual feature selection and rigorous cross-dataset evaluation for developing more reliable misinformation detection systems.

Future Internet doi: 10.3390/fi18110591

Authors: Tiago Bastos Cármen Guimarães Leonor Teixeira

Background: Organizations face increasing pressure to sustain economic performance while simultaneously addressing the demands of sustainability, resilience, and human-centered practices. Thus, this study analyzes how the maturity dimensions established for Industry 4.0 (I4.0) can facilitate the transition to Industry 5.0 (I5.0) by assessing the digital capabilities and workforce of Portuguese companies with systems capable of transitioning to smart factories. Methods: Shift4.0 survey data, based on an adaptation of the IMPULS model, were analyzed for 380 firms with complete responses to the smart factory and smart operations items. Principal axis factoring (PAF) with oblimin rotation was used to explore the underlying factor structure. Sectoral differences in factor scores were assessed using Welch’s ANOVA, followed by Games–Howell post hoc comparisons, while differences in workforce-related items were examined using the Kruskal–Wallis test, followed by Dunn–Bonferroni comparisons. Results: Two moderately correlated factors emerged: Operational Digital Integration and Automation, and Digital Infrastructure and IT Security. The first did not differ across sectors; the second and both workforce items (Employees Skill Sets and Skills Acquisition) did. Most firms scored at low maturity levels. Conclusions: The exploratory, cross-sectional findings are not a direct I5.0 readiness measure. They suggest that an item-level structure can reveal capability profiles concealed by minimum-score aggregation, and that infrastructure and workforce capabilities, potential I5.0 enablers, are unevenly distributed across sectors among responding firms with production operations.

Future Internet doi: 10.3390/fi18110590

Authors: Xinfeng Dou Xuexin Huang

Early warning of corporate financial distress matters to creditors, investors, and regulators, and automated monitoring systems re-score listed firms after every disclosure. Under China’s Special Treatment (ST) regime, most designations are announced with the annual report that conventional one-year (k → k + 1) designs use as input, so part of their accuracy comes from information released with the outcome. We therefore build the timing constraint into the label: year-k statements predict a firm’s first ST or *ST (delisting-risk warning) designation in year k + 2, eligibility uses only information available at scoring, and we audit the remaining leakage channels. Using 13,702 A-share firm-years (2019–2023), we make three contributions. First, on the primary cohort (k = 2023), whose model and thresholds were selectable at scoring, a random forest reaches an area under the receiver operating characteristic curve (AUC) of 0.856 (95% CI [0.811, 0.899]), and tuned logistic regression and gradient boosting perform similarly. Second, frozen validation-year thresholds (2%, 4%, 15%) produce four review tiers whose ST rates rise monotonically, from 0.63% to 33.33% on the primary cohort (pooled: 0.76–25.88%); at equal review capacity, one profitability ratio finds almost as many events, so the tiers add a probability scale, not a better ranking. Third, every score carries archived row-level SHapley Additive exPlanations (SHAP) contributions and a replayable tier assignment.

Future Internet doi: 10.3390/fi18110589

Authors: Leandro Pazmiño-Ortiz Alan Cuenca-Sánchez Ivonne Maldonado-Soliz

Tiny Machine Learning (TinyML) is expanding beyond compact convolutional and feed-forward models toward Transformer-family architectures adapted to microcontrollers and extreme-edge systems. This systematic review evaluates the deployment readiness of Tiny Transformers and closely related compact sequence models under severe memory, latency, energy, and runtime constraints, while examining security, privacy, reliability, uncertainty, reproducibility, sustainability, benchmarking, and application maturity. The review followed PRISMA 2020, and its methodological record was retrospectively registered on the Open Science Framework (OSF identifier 7fazq; review period: 2 March–1 June 2026). Searches across major bibliographic databases and supplementary sources identified 355 records; 213 unique records were screened, 87 full texts were assessed, and 81 reports were included in the qualitative synthesis. Direct Tiny Transformer deployment evidence remains limited and heterogeneous but converges on four complementary feasibility pathways: memory-aware scheduling, task-specific architectural downsizing, reduced-precision and runtime co-design, and sparse architecture–deployment co-design. Across the evidence base, support was substantially stronger for real-device execution and resource-aware optimization than for implemented security and privacy mechanisms, uncertainty calibration and out-of-distribution handling, reproducible deployment pipelines, standardized cross-platform benchmarking, system-level sustainability assessment, and longitudinal field validation. The evidence therefore supports a restrained conclusion: Tiny Transformers are becoming technically viable for selected constrained intelligent systems, but trustworthy, reproducible, and operationally validated deployment remains immature. Future research should prioritize cross-platform evaluation, transparent memory and energy reporting, quantified trustworthiness overhead, reproducible deployment artifacts, and field-oriented validation under realistic operational constraints.

Future Internet doi: 10.3390/fi18110588

Authors: Haoxing Zhang Zengrui Li Jiayi Guo

The rapid motion of Low Earth Orbit (LEO) satellites causes highly time-varying downlink quality, while the radio-frequency energy available during a single pass is strictly limited. Dynamic Programming (DP) optimizes discrete power allocation, but repeated optimization occupies processing capacity shared with other onboard tasks. This motivates compact policies that retain useful transmission performance with lower online computational cost. Hard-label imitation learning neither captures the value similarity among power actions nor guarantees the energy constraint. This paper proposes a lightweight online power-control method based on value-aware policy distillation. DP is used offline to generate complete state–action values, which are converted into soft supervision targets for training a compact neural network. During online inference, energy-feasible action masking removes power levels that exceed the remaining energy budget, thereby ensuring strict constraint satisfaction. The resulting policy observes the actual link and geometric states at each time slot and dynamically reallocates the remaining energy. Simulation results show near-DP average performance under nominal conditions, improvement over hard-label training, and adaptation to signal-to-noise ratio (SNR) forecast errors and along-track ephemeris errors. Rolling cross-level greedy achieves better transmission performance, while the proposed policy uses substantially less CPU time in the tested implementations, illustrating a performance–computation trade-off.

Future Internet doi: 10.3390/fi18110587

Authors: Jiang Xu Nigang Sun Yuanyi Zhang Haotian Zhu

The large-scale collection and circulation of passenger travel data in railway operations have dramatically heightened the risks of personal privacy leakage. In recent years, Decentralized Identity (DID) has become a vital component of privacy protection research for railway passengers by enhancing their autonomous control over identity information. However, existing DID-based railway passenger privacy protection schemes typically rely on static identifiers for continuous authentication within a single service domain. Consequently, different travel records can be linked into continuous spatiotemporal trajectories, directly exposing passengers’ real-world identities and resulting in severe privacy breaches. To address this issue, this paper proposes a trip-scoped dynamic DID-based privacy protection scheme for railway passengers. In this scheme, dynamic DID separates identity registration from the railway platform, and each DID’s lifecycle is strictly confined to a single trip, eliminating linkability through temporal correlation and behavioral pattern analysis. For selective disclosure at ticket purchase and gate verification, the scheme adopts the established BBS+ signature mechanism, allowing passengers to reveal only the minimum necessary verification information and achieving complete decoupling of identity authentication from travel trajectory data. Through security analysis and simulation experiments, the proposed scheme is demonstrated to be secure, effective, and feasible, offering a promising privacy protection solution for the railway industry.

Future Internet doi: 10.3390/fi18110586

Authors: Najem N. Sirhan Riyad Alrousan Sami I. Serhan Tariq Khraishi

Machine-learning models that predict short-horizon cellular throughput underpin adaptive video streaming, cloud gaming, and edge orchestration. On the widely used Raca et al. 5G drive-test dataset, such models can report low error. We show that, on this dataset, these headline numbers can be optimistic and, more importantly, that aggregate metrics conceal where and how the models fail. Using the saturated file-download sessions, where throughput reflects deliverable capacity rather than application idling, we (i) quantify the optimism of random-row evaluation relative to a deployment-realistic leave-sessions-out (LOSO) protocol, negligible at a 1 s horizon but reaching ∼32% at 8 s, and show that a trivial persistence baseline significantly beats the learned models at the two shortest horizons; (ii) show by ablation that most reported accuracy is carried by autoregressive throughput features rather than radio/context indicators and that the indicator-only mapping generalizes poorly across sessions; (iii) find, contrary to the intuition that handovers dominate error, that absolute error is lower around transitions and concentrated in the high-throughput regime (Gini 0.71), but that, in scale-invariant terms, the error is largest at low throughput and that transitions carry only a modest, noisy under-prediction tendency (event-aligned trough ≈−6 Mbps) that nearly vanishes under linear-space training (≈−1 Mbps), i.e., largely a log-space retransformation artifact; and (iv) find, using paired per-session bootstrap tests, that lightweight transition-context features give no significant global gain, a (mostly) null result. All headline numbers carry per-session bootstrap confidence intervals; the self-lag finding further replicates on an independent HSDPA drive-test dataset from Australia. We recommend leave-sessions-out evaluation, a persistence baseline, self-lag ablation, and scale-invariant, bias-aware reporting as a minimal honest protocol.

Future Internet doi: 10.3390/fi18110585

Authors: Mohammad Al-Omar Ayman Kamel Salim Jaradat Tariq Alomari Ala Mughaid Luay Tahat

Uncertainty-aware edge–cloud intrusion detection improves analytical efficiency by escalating ambiguous Industrial Internet of Things (IIoT) traffic for deeper cloud inspection, but this mechanism can itself become an attack surface when uncertainty is deliberately manipulated. This study formulates Adversarial Epistemic Load Amplification (AELA), a multi-stage attack in which protocol-valid and operationally plausible observations are optimized to increase epistemic uncertainty, selective escalation, cloud utilization, queue pressure, and analysis delay without relying on direct classifier evasion. Ten AELA variants are defined across uncertainty intensity, injection rate, stealth, persistence, and targeted intrusion activation. To counter this threat, an Epistemic Resilience Framework combines an Epistemic Escalation Risk Score (EERS) with an Epistemic Load Amplification Score (ELAS) and applies adaptive uncertainty gating, capacity-aware admission, source token control, redundancy suppression, local deep inspection, priority escalation, and closed-loop recovery. Experiments on the X-IIoTID dataset show 98.74% baseline accuracy and 0.996 AUROC. Under the strongest AELA condition, uncertainty increases by 98.5%, 78.6% of observations are escalated, peak cloud utilization reaches 1.46, queue pressure reaches 2.31, normalized delay reaches 2.07, and targeted intrusion success reaches 88.7%. The proposed EERS–ELAS defense detects 96.8% of AELA activity, reduces escalation by 57.9%, prevents 94.7% of saturation events, preserves 98.1% priority threat recall, and reduces targeted intrusion success to 9.4%. These findings show uncertainty must be governed with analytical capacity to preserve resilient edge–cloud intrusion detection under adversarial pressure.

Future Internet doi: 10.3390/fi18110584

Authors: Muborak Yakubova Maksat Amreev Gulnar Sadikova Muratbek Yermekbaev Bakytzhan Mustafina Nurzhigit Smailov

Failover planning for geographically redundant Multiprotocol Label Switching (MPLS) backbones must distinguish directional rate adequacy, packet deadlines, and interruption loss. This study develops a reproducible assessment combining G.729 traffic accounting, stochastic route/recovery trials, and independently verified finite buffer packet events. At 462 calls and 11.85 Mbit/s per direction, three-realization delivery by 100 ms averages 99.721% for independent phases and 90.818% for a 4 ms phase cluster. A matched comparison across 64 policy cells admits 30 using mean delay and blackout criteria, 19 using the full packet criterion, and eight using a sufficient all-phase bound. These counts persist at timely delivery targets of 98%, 99%, and 99.5% with the original blackout limit. Normalized analysis and five traffic levels from −20% to +20% identify when additional capacity, deadline headroom, or shorter recovery is required. A targeted 800-trace extension uses fifty realizations per cell and bounded one-sided release delays of 1–2 ms. The tested admission contrast persists, while shorter typical deadline miss runs can coexist with slightly worse aggregate timely delivery. The contribution is a verified comparison of conditional planning decisions and continuity consequences. OPNET illustrations provide configuration provenance; operational transfer requires calibrated traffic, service, and recovery inputs.

Future Internet doi: 10.3390/fi18110583

Authors: Ruize Guo Dianjun Xie Biaokai Zhu Xin Liu Yixue Geng Xiangbo Meng

Smart contract vulnerability detection is a critical challenge in the rapidly evolving blockchain ecosystem. Existing LLM-based auditing approaches face two major bottlenecks: (1) generated explanations often suffer from logical inconsistency and omission of key risk indicators, and (2) conventional preference optimization (e.g., DPO) performs only coarse sequence-level tuning, suffering from the inherent "lost-in-the-middle" bias that overlooks mid-sequence vulnerability-critical tokens. To address these issues, we propose TiDPO-SCVD, a token-importance-guided preference reinforcement framework for explainable smart contract vulnerability detection. Our method adopts a three-stage pipeline: continual pre-training on large-scale smart contract corpora to inject domain-specific security knowledge; multi-task supervised fine-tuning with a joint loss to simultaneously optimize detection accuracy and explanation completeness; and a hybrid token-weighting mechanism that integrates gradient attribution with Gaussian positional priors, combined with triplet regularization, to achieve fine-grained preference alignment at the token level. Extensive experiments on four prevalent vulnerability types, including reentrancy, timestamp dependency, integer overflow, and delegatecall exploits, show that TiDPO-SCVD consistently outperforms existing competitive methods across detection metrics. Moreover, both automatic LLM scoring and human expert evaluations confirm that our generated audit reports achieve superior performance in logical correctness, comprehensiveness, and interpretive clarity.

Future Internet doi: 10.3390/fi18110582

Authors: Khrystyna Zub Ivanna Dronyuk Oksana Mulesa

Web question answering often relies on passages that repeat the same source, offer independent support, or affect an answer only when considered together. We present PQ-EID, a method for examining how evidence interacts while accounting for its provenance. The method groups passages by claim and source information, evaluates evidence coalitions, and uses Möbius decomposition to identify individual and joint contributions. It also assesses whether an interaction keeps the same sign across plausible provenance groupings, subject to explicit bounds on evaluation error. Controlled experiments recovered known interactions and showed how mistaken source grouping can change their interpretation. A 10-question Question Answering with Conflicting Contexts (QACC) pilot revealed sensitivity to duplicated passages, although raw-context counting gave better predictive scores. We also evaluated two instruction-tuned language models on 60 QACC questions. Their coalition scores showed substantial higher-order interactions and sensitivity to the order of evidence. A separate controlled study of 80,000 trials examined when the conditional sign certificate succeeds and how it can fail if the true provenance structure is excluded. In the simulation with uncalibrated partition weights, 3.3–9.6% of the issued signs were wrong in the regimes where signs were issued. The real-data experiments use domain- and claim-based source proxies, without verified derivation chains. PQ-EID is therefore presented as a method for auditing evidence structure; its statistical guarantees apply under the stated structural and evaluation assumptions.

Future Internet doi: 10.3390/fi18110581

Authors: Viktoria Dmitrieva Nguyen Duc Tu Artem Volkov Ammar Muthanna Andrey Koucheryavy Dmitry Kozyrev János Sztrik

Future sixth-generation networks are expected to support immersive and telepresence services in which users interact with digital environments through wearable and multisensory terminals. In dense and mobile scenarios, static access support may become insufficient, since the radio-link quality and service availability can change as the user group moves. This paper investigates a robotic network coordinator that operates as a mobile base station for a group of users wearing telepresence suits. The coordinator position is linked to the communication performance through the signal-to-interference-plus-noise ratio, the achievable transmission rate, and an M/M/1 queueing model. To support dynamic positioning, we propose a two-stage FCM–ACO method. First, fuzzy c-means clustering identifies the current user groups and determines their cluster centers. Then, ant colony optimization constructs a feasible trajectory for the mobile base station under the movement constraint, using a QoS-oriented objective that reflects packet delay and radio-link quality. Numerical results compare FCM–ACO with FCM, K-means, and K-means–ACO in terms of BS trajectory length, coverage ratio, delay, cost, and computation time. FCM–ACO achieves the shortest BS trajectory among the four evaluated configurations, reducing the distance traveled by approximately one third relative to FCM, nearly one half relative to K-means, and approximately 9% relative to K-means–ACO. This reduction comes with tradeoffs in coverage, delay, and computation time.

Future Internet doi: 10.3390/fi18110580

Authors: Hang Song Jiepeng Tang Runwei Chen Jinghui Zhuang Meiyi Chen Yiming Xu

Future Industrial Internet systems are moving beyond site-bounded automation toward wide-area, distributed, and sustainability-constrained industrial intelligence. Connectivity, computation, and energy sustainability are still usually reviewed as separate design problems, even though industrial service chains must satisfy all three jointly. This survey examines service-chain placement under coupled connectivity, compute, and energy-sustainability constraints. It develops a three-continuum framework—connectivity, compute, and sustainability—and proposes six admissibility principles (P1 reachability, P2 criticality-preserving locality, P3 migration benefit, P4 continuity, P5 sustainability boundary, and P6 observability) as a falsifiable screening logic. Representative Industrial Internet of Things (IIoT) scenarios are used to examine how latency, reliability, coverage, workload locality, service continuity, and carbon constraints interact. The reviewed evidence suggests that satellite connectivity is admissible only when the workload is delay-tolerant, sparse, or resilience-oriented, and when local execution remains available for hard-real-time and safety functions; compute-continuum design requires workload stratification rather than indiscriminate offloading; and sustainability claims are meaningful only when energy and carbon are evaluated jointly with industrial service constraints. Because evidence maturity is asymmetric—terrestrial private networking and local edge computing are better supported than orbital execution and cross-domain carbon-aware orchestration—the six principles are proposed as rejection rules: a service is retained locally unless a candidate tier satisfies P1–P6 under the declared failure model and accounting boundary. Evidence is graded as conceptual/standards, simulation, prototype, testbed/pilot, or field deployment; claims are limited to the corresponding tier.

Future Internet doi: 10.3390/fi18110579

Authors: Olga Torstensson Dmytro Prokopovych-Tkachenko Valerii Lakhno Alona Desiatko Serhii Fedotov

Cloud integration can be economically attractive yet operationally unsuitable when dependent services exchange traffic over constrained network paths. This study develops a network-aware extension of a time-dependent, three-objective integration model and specifies a cognitive architecture for admitting, deferring, or escalating Pareto-based decisions. Strategic portfolio selection is separated from operational routing; infrastructure and service-dependency graphs connect both levels. The central mechanism evaluates whether the recommended action remains consistent across explicitly defined uncertainty scenarios, while independently checking service constraints, telemetry quality, authorization, and execution budgets. Numerical evidence is separated into an established synthetic portfolio benchmark, an independent reconstruction from its published inputs, and new deterministic network and admission-rule fixtures. Enumeration of the 18-system benchmark confirms 83,657 feasible portfolios and 446 nondominated solutions. Reanalysis of 30 paired run records preserves the original qualitative trade-off: a reference-direction NSGA-III-type baseline improves IGD+ and exact-front coverage, whereas NSGA-II uses less recorded CPU time. This comparison is a secondary reanalysis of offline optimizer records, not a measurement of real-time performance. A new three-service, two-route fixture contains 16 configurations, of which three satisfy the network-aware requirements; an additive compromise using the default route violates the specified end-to-end delay bound. A seeded stress test over 2547 synthetic instances, built from three service-dependency graph families on networks of two or three parallel logical routes between a local and a cloud domain, with traffic shifts, congestion, and route failures, reproduces this effect beyond the single fixture. A nine-scenario admission test yields recommendation agreement of 2/3, correctly withholding an automatic action at a threshold of 0.8, while a narrowed three-scenario set yields full agreement. These are deterministic model-based checks, not measured improvements in a deployed network; no trained predictor, emulation, or closed-loop controller is evaluated. The contribution is therefore formal and computational: an auditable connection between portfolio optimization, service dependencies, and decision-specific uncertainty. A separately identified evaluation protocol defines the learning, telemetry, ablation, and real-traffic evidence still required to establish closed-loop effectiveness.

Future Internet doi: 10.3390/fi18110578

Authors: Samia Kouki Firoz Khan Nithya Rekha Sivakumar R. Anbarasu Prabu Kaliyaperumal Sumendra Yogarayan Rajakumar Perumal

This study investigates whether learning behaviour in DDoS intrusion detection is determined solely by learning architecture or is also influenced by how the investigated network traffic is represented. A representation-aware experimental methodology is developed in which traffic representation is treated as the primary manipulated factor within each benchmark dataset, while benchmark harmonization, preprocessing, learning architectures, and evaluation protocols are controlled. Three complementary representations—Flow, Header, and Hybrid—are constructed from the same selected traffic instances within each benchmark dataset and evaluated using five representative machine learning and deep learning architectures. Predictive performance, experimental reproducibility, cross-representation stability, confidence calibration, and representation sensitivity are assessed using repeated experiments and statistical validation. The results show systematic differences across representations in predictive performance, cross-representation stability, calibration, and representation sensitivity under the controlled experimental conditions. These findings indicate that traffic representation is associated with differences in the discriminative structures made computationally available to learning algorithms rather than serving solely as a preprocessing mechanism. Based on these observations, the Representation-Dependent Discriminative Structure Principle (RDDSP) is proposed as a context-specific conceptual hypothesis describing how representation-exposed discriminative structures may interact with the inductive characteristics of learning architectures to shape learning behaviour in the investigated DDoS intrusion-detection setting. The empirical evaluation is restricted to a common attack space comprising SYN Flood and UDP Flood, together with Benign traffic, and uses a controlled, class-balanced evaluation distribution. Broader applicability of the findings and RDDSP remains subject to validation across additional attack categories, independently collected datasets, and operationally representative class-prevalence distributions.

Future Internet doi: 10.3390/fi18110577

Authors: Ibrahim Khormi Priyadarsi Nanda Mohammed Y. Aalsalem Manoranjan Mohanty Ibrahim Aqeel

Technological innovations have exposed vulnerabilities in traditional certificate verification systems. Conventional methods of verification cannot handle complex, evolving fraud systems and are susceptible to manipulation and human error. They can verify the authenticity of validly signed certificates but fail to detect behavioral fraud throughout the certificate lifecycle. This paper presents SmartGate—a secure framework that combines AI and blockchain to detect behavioral fraud and protect the integrity of academic certificates. The framework uses AI and feature engineering to capture user behavioral patterns indicative of fraud. The proposed SmartGate is designed to screen a certificate’s behavioral context before allowing it to be stored on the blockchain. University student records data were used to evaluate the proposed SmartGate. Specifically, the normal academic records and certificate workflow are real, while the suspicious cases were synthetically generated via controlled fraud scenarios. A machine learning algorithm is used to model the relationship between user behavior and the legitimacy of certificates. The framework introduces an SHAP explainable AI approach to interpret the outcomes for transparency and trust. To provide a secure, tamper-proof system, blockchain is deployed to anchor the approved certificates. A prototype of the proposed SmartGate has been developed. As a proof-of-concept implementation, the XGBoost model was used to validate the certificate-screening stage of the proposed SmartGate system. Also, a local Ethereum environment was used to implement the blockchain anchoring stage. Experimental results demonstrate the technical feasibility of the proposed SmartGate framework, including its ability to provide feature-level model explainability, within a controlled environment.

Future Internet doi: 10.3390/fi18100576

Authors: Alex Galis

Generative Artificial Intelligence (Generative Agentic AI) models, such as Large Language Models (LLMs), and agentic AI have recently gained widespread attention in both research and industry. They have rapidly emerged as a concept that elevates Generative Agentic AI from passive text generators to the next generation of agentic AI as autonomous, goal-driven agents capable of perception, planning, and dynamic action execution—Coherent Generative Agentic AI (CGA). In addition, CGA provides supervisory intelligence by maintaining persistent, evolving internal world models of the environment, networks, resources, agents, user intents, and policies. CGA reasons over time, runs long-horizon plans, checks consequences against its internal models, and coherently orchestrates lower-level controllers. It sits above network controllers, providing a global, semantic, model-driven brain. The paper starts with a review of CGA functional and non-functional characteristics in the context of key research and development results of agentic AI. It introduces a coherence-first architectural construct and system variable across all future 6G/7G networking planes (data-, transport-, management and operations,- orchestration,- autonomicity-, programmability-, and AI_as_a_Service-plane). It proposes a new deterministic CGA mathematical (CGAM) model that provides a new abstraction and theoretical foundation for analyzing the systemic consequences of different structural design and optimization options. It defines the relation between coherence and intelligence. The CGAM model represents the compute–communications continuum as a partially observable, intent-driven, contract-governed, multi-agent dynamical system with persistent memory and coherence constraints. We also included detailed CGAM validation and evaluation measures. This paper explores and details the CGAM’s role in 6G and 7G networking, as well as the challenges and opportunities for future work they present. We present the CGA-enabled 6G networking as well as 7G networking. Our investigation maps the current development landscape and points to future research and development in using Generative Agentic AI for networking.

Future Internet doi: 10.3390/fi18100575

Authors: Aksultan Mukhanbet Serik Aibagarov Yedil Nurakhov Timur Imankulov

Offline reinforcement learning offers a way to develop O-RAN controllers from historical network data without testing new actions on a live network. Its usefulness, however, depends on whether the logs preserve meaningful action sequences and whether policy values can be estimated reliably. We examine these questions using the public TRACTOR dataset and a preference-conditioned multi-objective Conservative Q-Learning (MO-CQL) controller. After checking timestamps and resource allocations, we retained 5438 transitions from 63,594 recorded action events and split them by experiment for training, validation, and testing. The recorded network state predicts throughput reasonably well, but the logged action adds little predictive information, and resource efficiency is difficult to predict. Comparisons with simple policies do not show a stable performance advantage for MO-CQL. The data also provide limited evidence for long-term evaluation: only 217 of 1422 adjacent test pairs form verified consecutive action sequences, and the longest such sequence contains four transitions. By contrast, a discount factor of 0.99 places weight on roughly 100 transitions. Estimates from fitted Q evaluation change substantially with training duration and the treatment of sequence boundaries. A separate experiment with known dynamics shows that splitting stored data into short files does not itself cause this problem. What matters is whether those boundaries change the return being estimated; neural approximation introduces further error. These findings offer a practical way to check the reliability of offline policy comparisons in O-RAN. Under a fixed FQE protocol, one MO-CQL checkpoint yields value estimates close to those of five separately trained scalar CQL policies while using fewer parameters and less storage. The efficiency objective is treated as a weak-signal, telemetry-sensitive proxy, and the main contribution is the evaluation-reliability framework.

Future Internet doi: 10.3390/fi18100574

Authors: Rohit K C Sridhar S K

The advancement of quantum computing poses a serious threat to conventional public key cryptography, as quantum algorithms could undermine the underlying mathematical problems used for security, potentially exposing private keys, enabling fraudulent digital signatures and unauthorized transactions, and compromising the confidentiality and integrity of protected data. The cryptographic mechanism plays a major role in achieving guaranteed ledger integrity and validity of a transaction in a blockchain. Nevertheless, commonly deployed signature schemes such as RSA, DSA, and ECDSA are susceptible to quantum-enabled attacks. We assess the viability of several quantum-safe cryptographic techniques for blockchain applications, including lattice-based, hash-based, and multivariate polynomial cryptography. Leading digital signature algorithms such as the Fourier-Lattice-Based Digital Signature Algorithm (Falcon), Module-Lattice-Based Digital Signature Algorithm (ML-DSA), Stateless Hash-Based Digital Signature Algorithm (SLH-DSA), and Module-Lattice-Based Key-Encapsulation Mechanism (ML-KEM) achieve strong security. However, existing multi-signature approaches focus on combining traditional cryptographic algorithms and post-quantum cryptographic algorithms to achieve security. This study evaluates Falcon-512 and ML-DSA-65 as two independent post-quantum signature components and combines them into a hybrid parallel dual-signature construction for blockchain-oriented transaction authentication. RSA-2048 and ECDSA P-256 are retained as classical reference schemes. The work has two distinct levels. The first is an empirical engineering study of key generation, signing, verification, transaction latency, throughput, signature size, storage, CPU usage, memory usage, tampering detection, and cross-scheme consistency. The second is a security extension motivated by a limitation of hybrid composition. Across the evaluated workloads, Falcon-512 achieved a throughput of approximately 283.88–290.28 transactions per second, whereas the parallel dual-signature construction achieved approximately 16.97–18.99 transactions per second. In the evaluated configurations, context-bound verification achieved zero cross-context acceptance and a 100% cross-context isolation rate. These findings demonstrate the importance of binding signatures to an explicitly defined security context and enforcing verification of both signature components. The security results provide a clearer distinction between legitimate and malicious transactions.

Future Internet doi: 10.3390/fi18100573

Authors: Qays Shaher Dani Merhi Bleik Ahmad Hani El Fawal Abbass Nasser Ali Mansour

We evaluate rolling-window XGBoost retraining (Adaptive XGBoost, AXGB) for leak detection on the simulated BattLeDIM benchmark, with an edge prototype and network simulation. Hourly supervised retraining on 6 h label windows (AXGB-S) reaches F1 = 0.993, but 99.1% of its refits repeat the last label; last-label persistence reaches 0.999. With a one-day label delay, AXGB-S (0.928), persistence (0.934) and an Adaptive Random Forest (0.931) detect leaks only after labels arrive. The shift defeating static XGBoost (F1 = 0.448) coincides with growing unlabelled background leakage, an important though not isolated explanation; metered demand decreased. A label-free, alarm-gated adaptive inflow model (AXGB-U) detected one of two held-out 2018 leaks after 1.2 h (0.36 false alarms per week) and two of four large abrupt 2019 leaks, after 0.8 and 12.8 h. An Arduino–Raspberry Pi prototype demonstrated the edge architecture once. In OMNeT++/INET simulations, injected loss reaches the detector almost one-to-one; sub-second delay has no effect. In a closed-loop channel model, AXGB-U detected the first test leak in every run up to 50% loss, but false alarms rose to 3 per week and four runs detected it after 27–28 h. Uncapped 1 Hz alarm bursts cost 53 times the 5 min cadence (30 min cap: 1.35 times).

Future Internet doi: 10.3390/fi18100572

Authors: Evelio Astaiza Hoyos Héctor Fabio Bermúdez-Orozco Nasly Cristina Rodriguez-Idrobo

Sixth-generation (6G) networks will serve large pre-trained artificial intelligence (AI) foundation models from cloud, edge and device tiers, but placing, compressing and scheduling them under latency, energy and accuracy constraints remains open. This paper proposes Massive AI Model Orchestration (MAIMO), a three-tier framework that manages model residency and request routing inside the radio access and core networks. The orchestration problem is formulated as a mixed-integer multi-objective program for which a Pareto-optimal solution is proved to exist and whose offline placement-routing version is shown to be NP-hard. An online controller couples a bidirectional long short-term memory (BiLSTM) traffic predictor with proximal policy optimization (PPO). The evaluation is simulation-only, in a purpose-built system-level simulator calibrated to standardized urban-macrocell channel models. Over 20 paired replications, MAIMO halves the mean end-to-end latency of cloud-only service (11.3 ms against 22.7 ms), serves ultra-reliable low-latency vehicle-to-everything requests in 2.1 ms on average with warm caches and attains the lowest service-level-agreement violation rate among schemes meeting the accuracy floor, while tying deep Q-network and contextual-bandit controllers on mean latency. A transparent analytic energy model gives 5.46 J per inference against 16.73 J for cloud-only service, and constrained carbon-aware workload shifting reduces operational carbon by 27.2%.

Future Internet doi: 10.3390/fi18100571

Authors: Efthimia Mavridou Michail Selvesakis Eleni Vrochidou George A. Papakostas

Recommender systems are a well-established area of the digital ecosystem supporting personalization in a wide range of domains in cloud computing environments. As recommendation functionality is increasingly delivered through cloud infrastructures, factors beyond recommendation accuracy, including latency and operational cost, have become critical for real-world applications. At the same time, Large Language Models (LLMs) have emerged as a promising direction for next-generation recommender systems. The goal of this work is to provide a comparative evaluation of different recommender systems based on trained conventional recommendation algorithms and zero-shot LLMs in a cloud environment with respect to recommendation quality, latency, operational cost and carbon footprint. A total of 18 different recommender systems are compared based on the MovieLens 100K dataset and a 20-item sampled-candidate setting, 13 of which utilize conventional recommendation algorithms while the remaining 5 are zero-shot LLMs. The results confirm the superiority of the neural collaborative model NeuMF and the embeddings-based matrix factorization method BPR-MF. In contrast, LLMs in a zero-shot setting underperformed while reporting a much higher estimated carbon footprint. Overall, this work provides an up-to-date benchmarking of trained conventional and zero-shot LLM-based recommender systems and highlights pivotal tradeoffs that should be considered when deploying recommendation systems in cloud environments.

Future Internet doi: 10.3390/fi18100570

Authors: Miklos Vincze Bela Molnar Miklos Kozlovszky

A patient’s imaging record is fragmented across modalities and viewers. CT and MRI arrive as DICOM volumes, whereas pathology is captured as gigapixel whole-slide images (WSI), each in its own coordinate system and scale. This paper presents a post-processing pipeline and augmented reality (AR) viewer that fuse separate CT, MRI, and WSI samples into a single common virtual space, so that effectively all a patient’s imaging data can be co-located and inspected together, simultaneously, within one interactive 3D space. Each modality is placed in a shared physical coordinate system through an externally supplied per-modality affine transform rather than by landmark registration. A central difficulty is the extreme resolution difference between millimeter-scale radiology and micrometer-scale histology. We bridge it without resampling the modalities onto a common grid, preserving each one’s native resolution, and we store the gigapixel volume as a sharded multiresolution pyramid whose native level is kept lossless for diagnostic use while the coarser overview levels are lossy, keeping the volume streamable without compromising the diagnostic layer. To render this span stably, we adapt the floating-origin principle from large-world rendering to the cross-scale medical setting as a focus-relative local coordinate frame that decouples the global scale span from GPU single-precision arithmetic and eliminates the positional jitter it would otherwise produce; we characterize this effect quantitatively, against a naive global frame, under progressive magnification. A single-pass GPU ray-marching traversal composites all modalities with independent per-modality controls (clip plane, opacity, intensity window), and a focus-and-context streaming strategy bounds on-device GPU memory for the visualization device. We evaluate conversion efficiency and fidelity, precision under extreme zoom, and native-resolution streaming performance on the headset against a fixed on-device memory budget.

Future Internet doi: 10.3390/fi18100569

Authors: Shafiq ur Rehman Muhammad Usman Khalid

Multi-LLM systems increasingly treat agreement across independently queried models as a check against hallucination. Because frontier models share training corpora, their errors can correlate, so independent models may converge on the same falsehood, turning agreement into a false consensus. We introduce SiGMA (Spectral Geometry of Multi-model Agreement), which reads the spectral and volumetric geometry of the per-question response cloud rather than agreement magnitude. Nine commercial LLMs were queried once each on TruthfulQA, SimpleQA, and FreshQA, and each question’s nine embedded answers were binned into five consensus states by the number of wrong answers. Mean pairwise agreement reads almost identically at total failure on all three datasets and cannot separate convergent from dispersed error. The volumetric descriptors flip sign. On adversarial-misconception items the all-wrong (AW) cloud contracts until its mode-spread volume is statistically non-separable from correct consensus (p=0.91, n=27), whereas it expands on the other two benchmarks. A tightly contracted cloud can therefore accompany total failure, so high consensus is not evidence of correctness. An exploratory query-type router lowers TruthfulQA AW false-accepts from 44% to 33%. Because total failure is rare, this regime rests on a small sample and is read as high-risk rather than as a prevalence estimate.

Future Internet doi: 10.3390/fi18100568

Authors: Kyrmyzy Taissariyeva Zhuldyz Kalpeyeva Yerlan Tashtay Zhansaya Ayapbergen

Indoor coverage infrastructure plays an important role in maintaining reliable LTE/5G services in shielded and structurally complex environments where outdoor macro-cell signals are strongly affected by penetration loss and internal attenuation. However, indoor coverage solutions are commonly evaluated using spatial measurements or short-duration before/after comparisons, while considerably less attention is given to their temporal interaction with the surrounding operational mobile network. This study presents a longitudinal operator-side assessment of LTE/5G network performance and stability during the successive activation of indoor coverage infrastructure. The experimental campaign was organized into three consecutive operational stages: a baseline state with both coverage-enhancement components inactive, a first intervention state following activation of the distributed indoor coverage component, and a final combined state after activation of the second localized coverage component. Temporally resolved operator OSS records were retained in chronological order and evaluated using a common staged framework. The analysis combined stage-wise descriptive statistics with segmented interrupted time-series regression, accounting for temporal trends, network-load variations, and serial dependence. The analysis focused on complementary network-side indicators representing the uplink radio environment, user-plane traffic dynamics, and 5G connectivity, including LTE and NR uplink interference, downlink and uplink PDCP volume, and the SgNB Add Success Rate. Across Stages 1–3, mean LTE uplink-interference levels were −108.20, −109.22, and −108.85 dBm, respectively, while the corresponding NR values were −110.35, −111.59, and −111.05 dBm. Although intervention-associated changes were identified in the statistical models, the baseline-derived persistence assessment detected no continuous adverse threshold exceedances lasting 24 h or longer during the post-baseline stages. PDCP traffic showed pronounced short-term fluctuations and progressively different stage-wise averages, demonstrating that network utilization changed during the observation period and providing essential temporal context for interpretation of the other KPIs. The mean SgNB Add Success Rate remained above 99.8% in all three stages and reached 99.92% in the final combined state, where its variability was also the lowest. Although transient KPI deviations occurred during the monitoring period, they did not develop into a consistent pattern of persistent deterioration across the monitored radio and connectivity indicators. The findings demonstrate the value of combining longitudinal operator-side KPI monitoring with segmented time-series analysis and explicit stability criteria. The study provides a structured assessment of staged indoor infrastructure deployment while recognizing the limitations of causal inference from single-site observational data.

Future Internet doi: 10.3390/fi18100567

Authors: Stefan Biševac Živko Bojović Petar D. Bojović

Multi-tenant Open vSwitch (OVS) deployments share a single datapath process and a common pool of central processing unit (CPU) cycles, creating risk of cross-tenant performance interference when sustained tenant traffic surges exhaust shared CPU cycles. We present a closed-loop optimal regulator that enforces per-tenant CPU budgets at the network interface. A two-state Linear-Quadratic Regulator (LQR), tracking CPU error and its backward-difference rate, is synthesized from an offline-identified first-order plant model and actuated via driver-level packet dropping within a single shared extended Berkeley Packet Filter (eBPF)/eXpress Data Path (XDP) program with per-namespace eBPF maps. The control loop executes in user space every 0.5 s; per-packet drop decisions execute in the driver at roughly 15–20 eBPF bytecode instructions. Evaluated on a gigabit testbed with single and dual OVS namespaces handling 512-byte User Datagram Protocol (UDP) traffic, the controller demonstrates tight set-point tracking: under active load (≥300 Mbps), single-namespace conditional mean tracking errors, in percentage points (pp), are +0.25 (3% target), +1.32 (4% target), and +0.18 (5% target); a dual-namespace pair holds the combined target within 1.9 pp with near-perfect fairness (J=0.9964). The three targets expose a clear throughput–CPU trade-off: the tight 3% budget is enforced by dropping roughly 60–63% of offered packets in total (commanded drops plus residual buffer loss), whereas the 5% budget delivers the highest throughput efficiency (up to 50.5% in dual-tenant operation). During a 100 Mbps → 500 Mbps step disturbance at a 5% CPU target, the controller peaks at 8.65% (+3.65 pp overshoot), with the initial transient resolving within ts≈2.8 s; no growing control-loop oscillations are observed. Against a single-point-tuned proportional–integral–derivative (PID) baseline at the 5% budget, LQR exhibits descriptively lower conditional mean tracking error (+0.18 pp vs. +0.77 pp, absolute difference of 0.59 pp; single-run point estimate, n=1, not statistically validated). Under 500 Mbps dual-tenant load, regulation saves 28.6% of host CPU cycles at a 4.1 pp throughput efficiency cost. The multi-tenant evaluation uses N=2 namespaces; control-law and datapath scalability to larger N is argued analytically by the O(1) per-tenant complexity, but physical-plant separability for N>2 requires separate empirical validation. The reference implementation, evaluation harness, and analysis scripts are publicly available.

Future Internet doi: 10.3390/fi18100566

Authors: Mahesh G.C. Md Mamunur Rashid Fariza Sabrina

A forward-secrecy claim protects past sessions only against the compromise it assumes. In lightweight Internet of Medical Things (IoMT) authentication, devices may stay in service for years, and a server may store more than its master key. We compare six published schemes at the source level and find that in five of them the published equations allow a key protecting past sessions to be recomputed from the state retained by the server or gateway. We then examine a recently published Ascon-based IoMT protocol in detail. Its forward-secrecy discussion considers only disclosure of the master key, yet an attacker who also obtains the registration records can recompute the transcript-protection keys and reconstruct an accepted historical session key. A ProVerif experiment, in which compromise is triggered only after the tested session completes, produces an executable trace of this reconstruction. The protocol also reuses Ascon nonces with fixed keys, contrary to the NIST nonce-uniqueness requirement, and its message flow provides no reverse key confirmation. An authenticated ephemeral Diffie–Hellman or idealised key-encapsulation contribution blocks the reconstruction in the symbolic models, and a further variant, S4, adds reverse confirmation and transcript-identity hardening. All results are symbolic, and X25519 and ML-KEM-512 serve only as deployment and size references.

Future Internet doi: 10.3390/fi18100565

Authors: Hugo Vega-Huerta Renzo Madge-Morales Alex Vela-Gutti Mario Chauca Adegundo Camara-Figueroa Oscar Benito Pacheco Rubén Gil-Calvo William Enrique Jorge Pantoja-Collantes Antonio Luna Figueroa Juan Carlos Lázaro-Guillermo Jessy Isabel Vargas-Flores

Internet Exchange Points (IXPs) constitute critical infrastructure for reducing dependence on external transit and improving network performance; however, their deployment in rural areas of Latin America and the Caribbean (LAC) remains limited, and the adoption of routing security mechanisms is incomplete. This study evaluates three dimensions: the regional evolution of IXPs, observable routing-security evidence among entities associated with IXPs, and the operational performance of a pilot deployment in Junín, Peru. The methodology combined a bibliographic analysis of sources indexed in the ACM Digital Library, IEEE Xplore, and Scopus with data from PeeringDB, Regional Internet Registry (RIR) delegation files, BGP tables, Looking Glass servers, RPKI validation, and MANRS records. For the Junín pilot, a before/after comparison was conducted during March 2024. Between 2016 and 2021, the number of active IXPs in LAC increased from 68 to 98 sites (+44%), while 60 IXPs met the inclusion criteria for the routing-security analysis in the available January–March 2023 extraction records; of these, 30 (50%) showed public evidence of at least one routing-security mechanism among their associated entities. In Junín, the observed mean RTT changed from 137 to 6.9 ms for the Meta measurements and from 138 to 43.8 ms for the Google measurements; because different endpoints were measured before and after deployment, these values represent observed route-specific changes rather than strictly comparable reductions and they do not constitute an isolated causal estimate of the peering effect because the design did not include a control site and, in Meta’s case, the regional cache was part of the same intervention. Netflix was assessed using traceroute because an equivalent ICMP measurement could not be obtained and, therefore, its results were not combined with the ping-based ones. The urban–rural comparison of transit costs relies on approximate values for Lima and Junín compiled from secondary information and pilot documentation, whose original official source could not be traced; it is presented only as an illustrative order of magnitude, not as a verified official or regional billing measurement. Taken together, the results provide site-specific empirical evidence from a rural pilot and a regional framework for studying traffic localization, routing security, and connectivity gaps while avoiding generalizing the site’s results to rural LAC as a whole.

Future Internet doi: 10.3390/fi18100564

Authors: Emad T. Elkabbash Alshaimaa Abo-Alian Tamer Abdelkader Mohamed Hashem

Differentially private stochastic gradient descent (DP-SGD) enables neural network training with formal privacy guarantees, but its utility depends on model architecture, data modality, and privacy budget. DP-RVFL freezes random hidden features and privately trains an output layer with a direct input connection. We compare it with Dense/CNN models, no-link and parameter-matched frozen controls, logistic regression and published DP-ELM. MNIST, CIFAR-10, Cardio and Heart are evaluated at ε = 0.1–8 using five training seeds. Gaussian gradient models use Poisson sampling, Adam, δ = 10−5 and add/remove-one adjacency. Primary ε = 2 accuracies are 90.30%, 38.61%, 72.99% and 63.48%, respectively. CIFAR-10 training takes 71.29 s versus 529.76 s for DP-CNN using eight CPU threads. DP-RVFL exceeds the no-link mean accuracy in 20 of 24 primary conditions, but the matched frozen control in only eight. Two additional executions preserve the same configuration and are reported separately. Five-pair exact tests do not establish statistical superiority. Diagnostics do not substantiate a reduced-clipping explanation, and Heart results vary across small partitions. The evidence supports lower training time relative to the tested Dense/CNN models; privacy budgets apply per model conditional on public benchmark preparation.

Future Internet doi: 10.3390/fi18100563

Authors: Khubaib Ahmed Qureshi Muhammad Sabih Muhammad Shoaib Siddiqui

Microblogs provide a continuously evolving source of information, but their short, noisy, contextual, and rapidly changing nature makes effective search, exploration, summarisation, and analysis challenging. This paper presents a unified framework for microblog intelligence that integrates credibility-oriented processing, semantic retrieval, faceted exploration, continuous extractive summarisation, topic-evolution analysis, media aggregation, sentiment analysis, and multidimensional analytics within a common architecture. The framework is organised into three functional layers: Credibility-Oriented Processing, Intelligence and Retrieval, and Analytical Intelligence. These layers are supported by a shared processing and representation flow and guided by five design principles: Unified Representation, Credibility-Oriented Processing, Progressive Intelligence, Temporal Intelligence, and Non-Generative Traceability. Facet-Value Pairs (FVPs) provide a structured representation that can be reused across retrieval, faceted exploration, summarisation, and analytical functions. The framework uses bot or suspicious-account detection and provenance information as early credibility-oriented signals; these signals provide filtering and source context but do not establish the factual truth of individual microblog posts. The framework also adopts extractive rather than generative summarisation so that summary content remains directly traceable to the underlying microblog posts. An interface-based case study demonstrates how the integrated architecture supports topic search, faceted refinement, temporal and geographical exploration, sentiment analysis, entity analysis, media exploration, and continuous summarisation. The contribution of the work is primarily architectural and integrative: it coordinates complementary microblog-processing capabilities through shared representations and a defined processing flow. A systematic quantitative evaluation of retrieval effectiveness, credibility filtering, summarisation quality, computational performance, and component-level contribution is identified as future work.

Future Internet doi: 10.3390/fi18100562

Authors: Roberto Pazzi Davide Facheris Davide Tosi

Digital Twins used as security replicas of cyber-physical systems must detect attacks while keeping decisions inspectable. Atop an existing entropy-gated neurosymbolic operator, we introduce an Entropy Analysis Layer (EAL), which treats estimator, window and stride as deployment variables with a two-sided gate |ΔHt|>τs calibrated per configuration, and a Semantic Explainability Layer (SEL), which exposes operator reasoning through three canonical SPARQL queries on a threat-intelligence-aligned ontology. Evidence types are separated: a 30-seed synthetic factorial; the official CIC-IDS2017 Friday capture, per second, with causal calibration and 4.46 h out-of-sample benign traffic; a minute-level robustness check; a synthetic stress test. On the capture’s single DDoS event, the two-sided permutation-entropy gate detects it in all nine overlapping, non-independent window/stride configurations (median delay 170 s, 33 false positives per benign hour with μ+3σ; 326 s and 2.7 per hour with the empirical 99.7% quantile); the one-sided gate detects it in none with μ+3σ and, via a single onset alarm, in one with the quantile. A rolling packet-rate z-score detects it after 43 s, with 91 false positives per hour. Shannon results depend on binning and calibration; no estimator dominates. Canonical queries run in 22–111 ms (median) at 5×104 triples. The layers provide auditable configuration evidence, not benchmark-level detection performance.

Future Internet doi: 10.3390/fi18100561

Authors: Jim Prentzas Ioannis Hatzilygeroudis

Generative Artificial Intelligence (AI) has increased in popularity in recent years, providing enhanced experiences to users and advancing many fields [...]

Future Internet doi: 10.3390/fi18100560

Authors: Oriol Agost Josep Rius Jordi Vilaplana Jordi Mateo

Representation compression alone does not determine serialized traffic in partitioned graph neural network (GNN) inference. Deduplication, partition topology, node identifiers, and batch framing also affect the wire size. We measure these effects for a one-layer graph attention network that transmits task-trained embeddings. Validation selected 64-dimensional embeddings across five seeds on Cora, CiteSeer, PubMed, and Flickr. These embeddings retained 97.9%, 97.7%, and 99.2% of full-feature accuracy; Flickr achieved 102.4% of full-feature accuracy. Under four balanced-random partitions, their single-round communication reduction factors relative to a one-time centralized float32 upload were 12.69×, 41.30×, 5.19×, and 3.01×. The predefined criterion of 95% accuracy retention and a minimum fourfold single-round communication reduction factor therefore passed on three datasets. Int8 transmission reduced the payload by a factor of four, but reduced wire traffic by only 3.66–3.67× because identifiers and framing remained and each nonempty int8 batch added a 4-byte scale. A centralized int8-upload comparator preserved full-feature accuracy within 0.0006. Relative to this comparator, float32 collaboration achieved single-round communication reduction factors of 3.19×, 10.34×, 1.31×, and 0.76×, with 0.76× indicating increased traffic on Flickr. A post-review balanced METIS baseline reduced four-partition serialized traffic relative to random placement by 89.6%, 97.6%, 88.9%, and 44.7%, respectively, under three placement seeds. Constant-width depth extrapolation reduced the four-layer communication factors to 3.17×, 10.33×, 1.30×, and 0.75×; deeper-model accuracy was not measured. This study uses partition-aware emulation on one workstation and does not measure physical devices or energy. Its main contribution is a reproducible wire-accounting protocol that distinguishes representation size, payload size, serialized traffic, host cost, and modeled link delay.

Future Internet doi: 10.3390/fi18100559

Authors: Abdulmajeed Albaqami Qazi Emad Ul Haq

Vessels disabling their Automatic Identification System (AIS) transponder evade AIS-based analytics. We present a framework for maritime dark-vessel screening—auditable candidate attributions for analyst review—coupling synthetic aperture radar (SAR) detection, AIS–SAR association, open-source intelligence (OSINT), and social network analysis (SNA). A Dynamic Kinematic Gate (DKG) grows its radius with speed and AIS staleness, up to 50 km, and demotes stale unmatched AIS to a low-confidence class. On 200 synthetic scenes, it raises true-match recall from 0.65 (capped static gate) to 0.88 and cuts spoofing alerts 4.1-fold via demotion; a fixed 10 km gate reaches 0.91 with fewer false confirmations, so the ceiling needs re-tuning. Re-benchmarked on five seeded SSDD folds with one evaluation code, YOLOv8s, YOLOv10s, and YOLOv11n reach a ship-class average precision (AP@0.50) of 0.940–0.955, on par with the deployed detector SLA-Net (Faster R-CNN with mask-attention and transformer modules; 0.938–0.941), whose two modules bring no significant gain. Without retraining, SLA-Net fails on inshore scenes of two other datasets (AP@0.50 ≤0.19). On a 12-day operational corpus the static-gate pipeline logs 3190 confirmed matches and 1606 dark detections; OSINT registry documents agree with the AIS-declared name for 90.1% and the flag for 71.8% of 2000 sampled vessels (consistency, not accuracy); a real one-month SNA graph shows no distinct bridge vessel. Attribution remains an analyst-adjudicated judgment.

Future Internet doi: 10.3390/fi18100558

Authors: Umair Khalique Huma Jamshed Urooj Waheed Fatima Ahmad Hussain

Future Internet ecosystems increasingly rely on distributed and intelligent data infrastructures, yet evidence of their evolution is fragmented across scholarly publications, patents, open-source repositories, and technology news. This study proposes the Semantic–Temporal–Authority Evidence Fusion Model (STA-EFM), a multi-source information-processing framework for identifying, ranking, and forecasting emerging data-infrastructure trajectories relevant to Future Internet ecosystems. STA-EFM integrates arXiv, PatentsView/USPTO, and GitHub records, with GDELT providing a limited complementary technology-news signal, harmonizes heterogeneous metadata, and assigns records to ten predefined technology classes using Sentence-BERT prototype similarity. A manually reviewed sample of 500 records was used to validate semantic relevance and taxonomy assignment. Source-normalized semantic evidence was combined with temporal activity, growth, momentum, recency, and source-specific authority indicators to generate monthly STA scores. A semantic threshold of 0.60 achieved an accuracy of 0.8800 and an F1-score of 0.9088, while taxonomy assignment achieved 0.8830 accuracy. Source-aware filtering reduced 2,069,979 records to 3337 relevant records. Federated Data Architectures achieved the highest mean STA score 0.3970, followed by Digital Twin Infrastructure and Graph and Vector Databases. Across the tested STA-level weighting configurations, Spearman rank correlations with the baseline ranged from 0.8667 to 0.9879. The framework provides an interpretable and traceable mechanism for monitoring emerging Future Internet infrastructure technologies.

Future Internet doi: 10.3390/fi18100557

Authors: Sara Lazzaro Carmen Licciardi Francesco Buccafurri

The increasing adoption of Internet of Things (IoT) systems in machine-to-machine (M2M) scenarios requires authentication mechanisms that support trust and accountability across organizational boundaries. Standard X.509-based mTLS mechanisms provide strong technical device authentication, while operational information such as the responsible legal entity, device role, authorization scope, and deployment context can evolve independently of the technical identity. This paper proposes a two-layer architecture that separates a relatively stable X.509 technical credential from a dynamic operational Verifiable Credential bound to the authenticated device identity. The base credential represents the device–entity association, while the context-aware credential adds a versioned domain profile, operational role, permitted operations, and deployment context. We define evidence-based issuance, lifecycle, runtime verification, and independent status mechanisms and relate the trust model to eIDAS concepts for certificate-based trust and electronic attestations of attributes. We experimentally evaluated the verifier-side processing cost of the proposed context-aware model, together with credential presentation and technical and operational status checking. The results show that context-aware verification including request-to-operation mapping requires approximately 226 microseconds per request on average, with cryptographic signature verification representing the dominant processing cost.

Future Internet doi: 10.3390/fi18100556

Authors: Kvitoslava Obelovska Andrii Abziatov Rostyslav Liskevych Volodymyr Vira Oleh Liskevych

The paper proposes a mathematical approach to adaptive superframe reconfiguration in hybrid MAC protocols for wireless mobile robotic networks. The considered network uses a hybrid access mechanism that combines contention-based and contention-free transmission periods within a superframe. The proposed approach is based on a probabilistic model of transitions between three traffic-priority states: low, medium, and high. A system of differential equations is developed to describe the state transitions, and analytical expressions for the stationary probabilities of the traffic-priority states are obtained. The probability of the high-priority traffic state is proposed as a decision criterion for adaptive adjustment of the relative durations of superframe periods. The influence of transition intensities on the state probabilities is analyzed, including their individual and combined effects. Within the examined range, the proposed probabilistic criterion is more sensitive to changes in the selected transition intensities at their lower values and tends toward saturation as these intensities increase. The proposed mathematical model provides a theoretical basis for adaptive resource allocation and superframe configuration in wireless robotic networks with dynamically changing traffic conditions.

Future Internet doi: 10.3390/fi18100555

Authors: Luigi Leonardi Gianni Antichi Gregorio Procissi Giuseppe Lettieri

Applications in virtual machines (VMs) rely on the guest Operating System (OS), which views only virtualized hardware, creating a semantic gap between this abstraction and the real execution managed by the hypervisor and host OS. Existing paravirtualization techniques bridge this gap through service-specific solutions, but implementing and extending these solutions can require changes to guest kernel components and host-side support, such as adding new hypercalls. We present a generic framework combining one-time paravirtualization during deployment with host-injected extended Berkeley Packet Filter (eBPF) programs, allowing different services to reuse the same guest infrastructure. After this initial setup, new services can be introduced by supplying additional eBPF programs, without modifying the shared guest infrastructure. Our prototype demonstrates how the same framework supports CPU pinning, dynamic memory expansion, and firewall offloading. In the evaluated benchmarks, these services achieve approximately 3.4-fold throughput with hyperthread-aware CPU pinning compared with execution without this mechanism and approximately 2.7-fold speedup with dynamic memory expansion compared with a fixed-memory configuration that resorts to swapping. Firewall offloading preserves legitimate traffic under a denial-of-service attack that prevents communication with guest-side filtering. These results demonstrate the framework’s applicability across different services; the contribution lies in their implementation through a common, extensible infrastructure.

Future Internet doi: 10.3390/fi18100554

Authors: Yang Li Yuhua Zhang Zhao Sun Yiyao Zhang

Autonomous mobile robots in dynamic indoor environments must manage future conflicts absent from current occupancy. We propose a time-consistent predictive navigation framework based on D* Lite and the Dynamic Window Approach (DWA). Its contribution is a unified temporal interface coordinating segmented prediction, continuous spatiotemporal validation, local risk handling, protected waiting, and time-matched D* Lite replanning. Assessment comprised 1026 core evaluations and a separate 90-case paired external benchmark, totaling 1116 runs in controlled numerical simulation. In the core paired comparison, navigation success was 97.8% for the full method and 34.4% for the current-occupancy fusion baseline. In the external benchmark, navigation success was 97.8% for the full method and 30.0% for a classical A*–DWA architecture. The full method had no collisions or missed physical-conflict detections, and the remaining 2.2% ended in active safe termination. Removing time-domain local risk handling reduced navigation success to 56.7%, with collisions in 10.0% of cases and missed physical-conflict detections in 6.7%. Across 180 single-factor perturbation cases, navigation success was 99.4%, and the remaining 0.6% ended in active safe termination. Within the tested scenarios, the coordinated mechanisms supported future-conflict handling and recovery from persistent blockages. Deadline violations remained, and all evidence comes from controlled deterministic simulation.

Future Internet doi: 10.3390/fi18100553

Authors: Yunkai Bao Ali Abbasi Khoa Nguyen Xin Wang Steve Drew

Federated learning (FL) trains models across geographically dispersed clients whose electricity carries very different carbon intensities. We present FedGreenCS, a carbon-aware approach to FL for IoT and next-generation networks. Clients are first grouped by data distribution similarity, using pairwise cosine similarity of model weights after a single local epoch. In each subsequent round, a sampled client whose carbon intensity exceeds a fixed tolerance offloads its training to the lowest-carbon available peer in its similarity cluster, while clients within the tolerance participate unchanged. Ordered dropout gives clients with a higher carbon intensity a narrower model. We derive a carbon cost model that separates computation and communication for each client and round, isolates the cost of the mandatory first-round broadcast, and states the design over the scaling rates and the tolerance. On a non-IID EMNIST partition with real regional carbon intensities, FedGreenCS with three carbon clusters reaches 70% test accuracy at 3.15 to 3.36 g CO2 of client-side carbon cost over three seeds, against 3.58 g for FedGreen and 15.12 g for FedAvg. With one model width for all clients, the offloading cuts this cost by 23% to 58%, but it raises the cost with two carbon clusters and always needs more rounds.

Future Internet doi: 10.3390/fi18100552

Authors: Ziyang Huang Lin He Jie Wan Ruirui Liu Weiqiang Tan Jing Sun Xiaoxing Cao

Automating the transition from heterogeneous security alerts to deception actions requires both decision constraints and verifiable execution. We present Constraint-Aware Deception Orchestration (CADO), a control-plane specification connecting normalized security events to typed deception action transactions with explicit scope, lifetime, verification, and compensation. The policy anchors each decision to a fixed consensus action and admits a one-level residual only after evidence, novelty, escalation-confidence, and model-based value checks. In a discrete-event simulation with three zones, four development attacker profiles, and 30 paired seeds, mean defensive utility was 0.7020 (95% bootstrap confidence interval: 0.7003–0.7038), versus 0.6953 for consensus; redirection precision increased from 0.6978 to 0.7282. The observed unsafe-action rate was 0.000046 for both policies. A held-out profile produced 98.68% consensus fallback, with small utility and precision gains accompanied by higher resource cost. An exploratory study of 26 parameter configurations identified sensitivity to admission thresholds and uncertainty penalties; safety-rate equality did not hold throughout the grid. A separate three-process HTTP prototype exercised 300 transaction tests, including explicit detection of 30 persistent rollback failures. These simulation and local integration results do not establish production effectiveness, calibrated safety guarantees, or interoperability with commercial H4 devices.

Future Internet doi: 10.3390/fi18100551

Authors: Shuyun Fang Junling Wang Xinxin Shao Derui Qi Fukun Chen

Multimodal Recommender Systems (MRSs) leverage heterogeneous data to alleviate data sparsity, yet they remain vulnerable to spurious correlations and modality noise that obscure the rationale behind recommendations. To address this, we propose the Multimodal Counterfactual Fusion with the Causality-Guided Contrastive Learning (MCFC) framework. MCFC integrates three core modules: a Multimodal Structural Causal Model that motivates the assumed causal paths of multimodal biases; a Counterfactual Feature Fusion module with causal effect-inspired weighting that estimates each modality’s contribution via score perturbations and suppresses noisy features; and a Causality-Guided Contrastive Learning mechanism that builds counterfactual in-batch negatives to enforce invariant and robust representations. Extensive experiments on four public Amazon datasets show that MCFC consistently outperforms state-of-the-art baselines on Recall@K and NDCG@K, improving Recall@20 by 29.9% over the strongest baseline on the highly sparse Baby dataset. Ablation studies confirm the contribution of each module, evidencing MCFC’s robustness to modality noise and strong adaptability in sparse interaction settings.

Future Internet doi: 10.3390/fi18100550

Authors: Eleftherios Varverakis Panagiotis Michael Demetris Trihinas

Zero-shot time series foundation models (TSFMs) forecast unseen series without target-specific training, offering a way to significantly reduce the deployment overhead in enterprise pipelines and industrial sensor streams. However, their performance is documented almost exclusively on pristine benchmarks, whereas industrial data streams feature missing values, sensor outages, and amplitude spikes. Whether these models absorb such faults natively, and thus whether costly imputation stages can be removed, remains unquantified. In this study, we evaluate the robustness of three TSFMs from distinct architectural families, Chronos (tokenization), TimesFM (patching), and TiRex (xLSTM), subjecting their context windows to random missing data, contiguous block missingness, and injection of point anomalies at escalating severities and assessing point accuracy (relative RMSE) and 80% interval reliability (PICP) against a clean baseline. The results show that random dropout and transient spikes are natively absorbed by all three models (accuracy loss below 3%), while sustained contiguous outages are the dominant failure mode, with patch-based TimesFM most exposed (up to 59%). TiRex stays well calibrated throughout, while tokenization-based Chronos yields persistently overconfident intervals. These patterns persist and the contiguous outage vulnerability is amplified on data with genuine operational anomalies. For the evaluated models, sensor streams, and tested corruption severities, the results indicate that direct inference remains close to the clean-context baseline under independent random loss and sparse positive spikes, whereas prolonged outages warrant additional missing-data handling.

Future Internet doi: 10.3390/fi18100549

Authors: Wisam Makki Alwash Weam Husham Aljabbari Belal Al-Khateeb Hasan Hüseyin Balık

Electronic Product Code Information Services (EPCIS) standardizes inter-organizational supply-chain visibility events, yet structurally conformant histories can contain contradictory lifecycle claims that undermine information integrity. We present EPCIS-MutBench, a controlled benchmark and layered validator for evaluating security-relevant semantic and lifecycle integrity in EPCIS 2.0.1 histories. Mutations are treated as integrity violations independently of accidental or adversarial origin; the benchmark evaluates detection and localization rather than attacker attribution. The evaluation uses 2000 deterministic clean histories and 20,000 primary mutants, including 16,000 single-fault and 4000 compound cases, with validator-independent typed ground truth. Across all primary mutants, JavaScript Object Notation (JSON) Schema plus compatibility-projected GS1 Shapes Constraint Language (SHACL) detected 18.27%, whereas the full event-local and role-aware lifecycle validator detected 95.29%. Among 13,429 single-fault mutants that passed both structural baselines, the full validator achieved 92.98% recall, with end-to-end intended-rule, event, and field/edge localization of 92.90%, 92.19%, and 92.14%, respectively. No false positives occurred on the 2000 generated-clean controls; however, in a separate sensitivity analysis that provisionally treated all 37 parseable external-source fixtures as clean, 19 were marked invalid. A separate attestation layer used cryptographic hashes, Merkle commitments, a digital signature, and public-testnet receipts to make post-freeze evidence modification detectable. The results demonstrate controlled, reproducible evaluation of EPCIS information-integrity assurance beyond structural conformance.

Future Internet doi: 10.3390/fi18100548

Authors: Firdous Kausar Lei Qian Asmah Muallem Sajid Hussain

Autonomous clinical agents increasingly act on Internet-of-Medical-Things (IoMT) endpoints, yet existing access-management mechanisms are human-centric, quantum-vulnerable, and poorly suited to continuous action-level enforcement. We present PQ-MedTrust, which couples an identity–behavior–context trust score with risk-dependent decay to post-quantum revocation of agent credentials based on Decentralized Identifiers and Verifiable Credentials, and we characterize where behavioral trust succeeds, becomes detection-latency limited, or fails. In a liboqs testbed on a general-purpose CPU, classical, hybrid, and PQC-only handshakes stay below 1.5 ms, but a hybrid authorized action (handshake, credential, delegation token, and request) occupies 21.1 KB against 1.8 KB classically. The trust-to-revocation path, which signs and pushes a revocation record, takes 0.07 ms with classical and 11.1 ms with SLH-DSA signatures. Over 1000 simulated runs per scenario, policy enforcement revokes every out-of-scope and one-shot mimicry attack (scope is a generator label, not a dose-range check), whereas adaptive stealth is detected in 20.2% of runs at 24 h and 75.3% at 120 h; without the generator’s risk labels, 24 h detection falls to 5.6%. The pooled benign false-positive rate is 1.8% under a reliable context but rises sharply when context degrades or benign agents emit the event class that marks adaptive stealth. A graduated response revokes none of 500 unannounced-emergency runs, although 92.4% reach clinician confirmation, and still revokes every policy violation.

Future Internet doi: 10.3390/fi18100547

Authors: Jun Zhang Dong Zhu Aiping Li Chuanyi Liu Zhaoquan Gu Binxing Fang

Honeypots and lightweight honeypoints record scanning, reconnaissance, and exploitation attempts against deliberately exposed services. In operational exports, however, observations from the same source are often split across dates, decoys, alerts, threat intelligence fields, and historical records. Ranking sources only by interaction count shows activity volume but not where a source appeared, when it returned, or which contextual signals accompanied it. We present Mizong, a source-level analysis protocol that converts these fragmented records into auditable profiles for longitudinal and cross-site analysis. Mizong links all records belonging to the same source and characterizes each resulting trace by interaction intensity, honeypoint breadth, observed date persistence, alert linkage, threat intelligence context, and historical activity. Each dimension has an explicit aggregation rule; missing and conflicting states remain visible, and every profile retains its contributing rows. We evaluate source profiles using 20 detailed Shenzhen-side exports containing 10,257 sources and 3,124,252 interactions from the 2025 National Games and Para Games assurance task. A top 5% interaction count queue covered 5.4% of multi-point sources, 14.5% of multi-export sources, and 2.4% of alert-linked sources. Source-set comparison found 778 identifiers in both assurance phases; a separate 76-day operational summary reported 603 identifiers at all three protected sites. These results show that activity volume and the other profile dimensions answer different analytical questions. Mizong provides a reproducible way to move from disconnected export rows to source profiles that can be inspected across time, phases, and sites.

Future Internet doi: 10.3390/fi18100546

Authors: Abraham George Mir Wajahat Hussain

The integration of next-generation immersive applications and ubiquitous computing calls for a fundamental shift in wireless networking to accommodate the massive, time-sensitive data transfers necessary for human–machine interactions. While emerging 6G technologies focus primarily on physical layer gains, existing siloed layered architectures remain a bottleneck by failing to leverage network diversity effectively. This paper introduces an augmented cross-layer architectural framework featuring a physical abstraction layer (PAL) and a cross-layer fabric (CLF) to decouple network logic from underlying hardware components. By using lightweight open interfaces, this model resolves cross-layer information asymmetry and enables interface-agnostic data distribution across heterogeneous wireless links. To evaluate operational feasibility, we formulate an open Jackson queuing model to quantify performance gains across aggregate throughput, packet latency, and server utilization. The analytical results demonstrate that while multi-interface aggregation significantly improves throughput and reduces latency compared to single-interface configurations, performance scaling is bounded by diminishing marginal returns due to physical layer scheduling overheads. This work establishes an analytical framework to quantify the performance bounds of wireless heterogeneity in future communication networks.

Future Internet doi: 10.3390/fi18100545

Authors: Seyed Salar Sefati Atanas Vlahov Razvan Craciunescu Bahman Arasteh Vladimir Poulkov

Internet of Things (IoT) service choreographies integrate services across devices, edge servers and the cloud without a central orchestrator. Existing quality-of-service (QoS)-aware composition methods use single-value QoS, separable objectives and one-shot binding, overlooking evolving service conditions, peer-link dependencies and rebinding costs. This paper proposes MDP-BWO, a Markov decision process (MDP)-guided Black Widow Optimization (BWO) framework that integrates state forecasting, dependency-aware binding and runtime adaptation. A three-state Markov chain learned from monitoring traces forecasts each service’s state occupancy and guides BWO over constrained, non-separable bindings; each participant rebinds locally by table lookup of a finite-horizon MDP policy over a value-selected shortlist. In simulations on synthetic and WS-DREAM-based instances (measured QoS, synthetic state dynamics; 14 configurations × 30 held-out seeds), MDP-BWO outperforms six baselines with reactive fail-over in every configuration, raising executed utility by 0.086–0.139 and the solved-request rate by 0.087–0.130 (Holm-corrected p < 0.05); latency on WS-DREAM and cost on synthetic instances do not improve over three of the population baselines. Most of the gain comes from MDP rebinding; with the same forecast and policy, MDP-BWO still outperforms five population-based searches in 11–14 configurations and is matched or slightly exceeded only by an exact mixed-integer search with 14–43 times longer mean planning time.

Future Internet doi: 10.3390/fi18100543

Authors: Denys Tsomenko Taras Panchenko

Edge databases must coordinate replica freshness, consistency requirements, and governance restrictions despite intermittent connectivity. This study specifies an Edge-Aware Database Management System (EA-DBMS) architecture centered on the Policy-Aware Distributed Edge Catalog (PADEC). Its architectural novelty within the reviewed literature is a shared metadata interface connecting these constraints to operator placement, local transaction admission, and reintegration. This design provides a basis for keeping permitted operations local during outages and combining authorized results across sites under explicit contracts. A worked example demonstrates how joint filtering rejects prohibited and stale candidates. The analysis formalizes policy-constrained operator placement and distinguishes NP-hard general cases from three subclasses admitting polynomial-time exact solutions. A reproducible analytical evaluation derives conditional latency trade-offs from published measurements, documented proxies, and explicit assumptions. At a 10 ms proxy catalog cost, local reads benefit in one of four paired network scenarios at the low end of the published execution-time range. Under baseline assumptions, reductive aggregation becomes beneficial above approximately 401 kB per window at a reduction factor of 0.189. In the modeled mixture, global coordination reduces the mean-latency advantage. The design targets industrial telemetry, disconnected local processing, and data-residency-constrained analytics, with catalog lookup cost and matched edge–cloud execution times as priorities for empirical evaluation.

Future Internet doi: 10.3390/fi18100544

Authors: Ihor O. Serhiienko Volodymyr O. Artemchuk

Cross-silo deepfake speech detection must accommodate organizations that observe different attack profiles and channel conditions while avoiding routine centralization of raw speech. We study federated learning (FL) for a compact detector combining linear-frequency cepstral coefficients (LFCCs) with frozen WavLM representations over heterogeneous partitions derived from ASVspoof 2019 and 2021. An official-metadata audit of the historical seed-42 Client-C split reproduces the reported 2572/3655 (70.37%) bona fide source overlap; a stricter unified-source audit against the complete A+B+C fit pool finds still greater dependence. We therefore repeat the principal audio evaluation with five source-disjoint splits whose holdouts are independently verified against the complete fit pool. Mean equal error rate (EER) is 3.836% for pooled centralized training, 4.622% for Federated Averaging (FedAvg), 4.095% for local C, 21.883% for local A, and 33.398% for local B. On the common Client-C holdout, A-only and B-only models transfer poorly, while FedAvg is on average 0.787 percentage points worse than pooled training and 0.527 points worse than local C; these comparisons do not establish native-domain benefit for A or B. Separately, a 20-seed controlled synthetic-embedding stress test examines integrity mechanisms under malicious participation. Increasing untargeted label-flip compromise produces progressively more negative benign–malicious update alignment and stronger aggregate-step attenuation; the signature persists across four classifier topologies and FedProx settings. Additional tests cover targeted label poisoning, a controlled trigger backdoor, zero-update free riding, and coordinated sign-flip model poisoning. We frame the contribution as an auditable experimental protocol and a bounded set of observations, not as a new detector, optimizer, attack, aggregator, or universal defense.

Future Internet doi: 10.3390/fi18100542

Authors: Geldi Xhafaj Umair B. Chaudhry Hamid Jahankhani

Spam and phishing email remain a persistent cybersecurity threat and a primary delivery vector for credential theft, malware, and ransomware. Detection accuracies above 99% are routinely reported. Such figures are typically obtained on small, single-source, monolingual corpora in which near-duplicate campaign templates span the training and test partitions. This study re-examines what they measure. A corpus of 99,707 emails is assembled from four independent public collections spanning multiple languages, campaign-aware deduplication is applied before splitting, and nine models—classical, convolutional, Transformer, and multimodal—are evaluated under two protocols: an in-distribution split and a cross-source split in which the evaluation source is withheld entirely from training. The multimodal model fine-tunes DistilBERT and fuses its contextual embeddings with thirteen structural features into a 781-dimensional representation classified by a multilayer perceptron. In distribution, six of the nine models fall within one F1 point of each other (ΔF1 = 0.0001, 95% CI [−0.0012, 0.0012], fusion vs. the fine-tuned text-only baseline). Against an identical classification head without the structural block—the controlled comparison isolating the structural contribution—fusion produces a small but statistically significant gain in distribution (ΔF1 = 0.0014, 95% CI [0.0005, 0.0023]; Δrecall = 0.0030, CI [0.0016, 0.0044]) that grows roughly five-and-a-half-fold under cross-source evaluation (ΔF1 = 0.0078, CI [0.0036, 0.0120]; Δrecall = 0.0107, CI [0.0031, 0.0180]). The contribution of multimodal fusion therefore scales with distribution shift rather than appearing only under it and is largest exactly where deployment conditions differ most from training. Against the strongest classical baseline (a linear SVM), the fusion model attains a false-positive rate less than one-fifth as large under cross-source evaluation (0.91% vs. 5.40%) alongside significantly higher F1 (0.9488 vs. 0.9188, non-overlapping bootstrap CIs); in distribution, it attains marginally higher AUC (0.9990 vs. 0.9985) at a marginally higher false-positive rate (0.94% vs. 0.37%). SHAP places 97.6% of signal in the text embeddings, yet that residual share is what produces the cross-source gain—attribution magnitude measured in distribution does not predict conditional contribution under shift.

Future Internet doi: 10.3390/fi18100541

Authors: Shafqat Hasan Umesh Biswas Thomas Benton Townsend Dimitrios Michael Manias Jared Saia Maxwell Young

A defense against distributed denial-of-service (DDoS) attacks is resource-competitive if the defense cost is small in the absence of attack and grows asymptotically more slowly than the attacker’s cost when an attack occurs. As the attacker’s cost grows, the system becomes increasingly more expensive to attack than to defend. Prior theoretical work shows that resource-competitive DDoS defenses are achievable under a simplified model. The primary contribution of this work is an empirical evaluation of whether these theoretically established resource advantages persist as the assumptions are progressively relaxed toward realistic operation. We evaluate the existing Linear and Linear-Power algorithms using controlled synthetic workloads, temporally ordered CIC-DDoS2019 traces, and a client–server implementation. Our results exhibit long-run behavior qualitatively consistent with the theoretical asymptotics, although the degree of agreement varies across attack types, and provide evidence that a resource advantage can persist under more realistic traffic. They also identify traffic structure and communication delay as important considerations in moving toward practical systems.

Future Internet doi: 10.3390/fi18100540

Authors: Ourania N. Kapsala Spiridoula V. Margariti Chrysostomos Stylios

Distributed computing environments, such as fog computing, are attracting increasing attention because they can bring data processing and storage closer to data sources and adapt to the ever-increasing demands of existing and emerging applications executed at the network edge. Workflow scheduling poses a significant challenge in these systems, hindering the full exploitation of their potential. However, workflow scheduling is an NP-hard problem. Despite the numerous existing approaches for addressing this problem, several limitations remain, including high execution latency, suboptimal scheduling performance, and inefficient utilization of the available computational resources. To address these challenges, this work proposes a novel hybrid metaheuristic mechanism, called EHHO-GWO, for workflow scheduling, with the objective of optimizing execution time, monetary cost, and energy consumption. Moreover, the proposed approach is comparatively evaluated against classical optimization algorithms, namely Genetic Algorithms (GAs) and Particle Swarm Optimization (PSO), as well as other recent metaheuristic approaches. Using FogWorkflowSim as the simulation framework and a set of scientific workflows as input, this study evaluates the effectiveness and efficiency of the proposed algorithm and demonstrates its superior performance in fog computing environments. The experimental results are statistically validated using the Wilcoxon signed-rank test, revealing consistent performance improvements over well-known competing scheduling algorithms.

Future Internet doi: 10.3390/fi18100539

Authors: Roman Banakh

The basis for this work is the growing prevalence of the Evil Twin attack on IEEE 802.11 networks, as it is the initial link in downgrade attacks that force the security protocol from WPA3 to WPA2 in transition mode, enabling handshake interception and an offline dictionary attack on the network password. For the study, a Wi-Fi access point and a device simulating the Evil Twin attack were deployed, and a multi-sensor Internet of Things (IoT) infrastructure was developed to record the received signal strength indicator (RSSI) from the access point. The resulting RSSI dataset was used to train machine learning models addressing three subtasks: attack detection, attacker localization, and antenna-type identification. Ten classification algorithms were compared using group-based nested cross-validation, keeping entire attack sessions together. Attack detection reached an accuracy of 98.8% with KNN, while the accuracy of localizing the attacking device among eight positions and identifying the antenna type was 74.3% (89.6% per attack session) and 36.3%, respectively, with KNN and XGBoost. The results demonstrate the feasibility of the proposed method as a first step towards reducing incident response time and narrowing the search area for the source of an Evil Twin attack on a Wi-Fi network.

Future Internet doi: 10.3390/fi18100538

Authors: Hari Hara Babu Saripalli Jyothsna Laxmi Saripalli Leonel Lagos Alexander Perez-Pons Himanshu Upadhyay

Networks of fixed, low-cost dose-rate sensors monitor radiological facilities, but the trustworthiness of their reports rests on heuristics: existing trust schemes score nodes by mutual voting or behavioral statistics, which fail against an adversary who falsifies data consistently with the physics the network can observe. This paper presents PRAHARI, a trust management framework in which an unmanned aerial system carrying a calibrated scintillator chain hovers at each node as a mobile metrological anchor. Trust evidence is a statistically principled compatibility test between two uncertainty-quantified measurements with an empirically verified false-alarm rate, feeding a multi-channel Beta reputation engine with wall-clock forgetting, endurance-constrained verification scheduling, and trust-gated dose mapping. Across 9640 paired Monte Carlo trials on a shielded-facility twin with randomized attack classes, PRAHARI is compared against three published network-internal trust methods, a physics-model residual detector, and an uncertainty-blind aerial band. Only aerial-reference evidence discriminates scaling and physics-consistent falsifiers (class AUC 0.80 and 0.75), while network-internal methods stay near chance (0.48 to 0.60), an advantage persisting across facility variants and up to half the adversary’s shielding knowledge, at a false-accusation rate one order of magnitude below every alternative of comparable ranking power. Sensitivity, ablation, reference-bias, and drift studies bound the claims.

Future Internet doi: 10.3390/fi18100536

Authors: M. Baqer

Abnormal operating conditions and equipment malfunctions in upstream oil and gas wells can compromise safety, disrupt production, and cause substantial economic losses. This study proposes FedRBF-SVM, a federated radial basis function (RBF) support vector machine (SVM) approach for collaborative anomaly classification across Industrial Internet of Things (IIoT)-enabled wells. Each participating well, represented as a client, trains and retains a local RBF-SVM using labeled samples and returns only a scalar decision-function score for each inference sample. The server averages these scores and applies a validation-calibrated decision threshold. The approach was evaluated on the 3W oil-well dataset in a simulated federated setting using different numbers of locally stored training patterns, three target false-positive rate (FPR) operating points, and a centralized RBF-SVM reference. FedRBF-SVM performance generally improved as the number of locally stored training patterns increased. With 2500 training patterns per client and a target FPR of 0.05, FedRBF-SVM achieved an F1-score of approximately 0.79, a precision of approximately 0.97, and a recall of approximately 0.67. These results demonstrate the feasibility of federated SVM for distributed oil-well monitoring while keeping training data and models at the clients.

Future Internet doi: 10.3390/fi18100537

Authors: Fernando Gutierrez Portela Oscar Augusto Diaz Triana Andres Felipe Yule Juan Camilo Galeano Bucurú Alexander Suarez Gómez Carlos Mario Paredes Diego Martínez-Castro

The expansion of the Internet of Things (IoT) has intensified cyber risk in Edge deployments with limited computational resources, where conventional intrusion detection systems are unfeasible. Aiming to reduce this gap, a lightweight anomaly detection system for IoT networks has been designed and implemented, structured through the CRISP-ML(Q) (Cross-Industry Standard Process for Machine Learning with Quality Assurance) lifecycle. Unsupervised models (Isolation Forest, OCSVM, K-Means, and Autoencoder) and supervised models (Decision Tree, Extra Trees, and Random Forest) were trained and compared on the BoT-IoT dataset, validating their operational behavior through real-time traffic captures during “TCP RST Flood”, “TCP SYN Flood”, “UDP Flood”, and “ICMP Flood” attacks on a Raspberry Pi 4B platform. Isolation Forest provided the most consistent balance between detection capability and computational cost, while the Autoencoder achieved detection rates of up to 98.8%. Among the supervised models, Random Forest achieved the highest ROC–AUC, reaching 98.5%, whereas Extra Trees showed the lowest inference time under attack conditions (0.50 s). The system was deployed on a Raspberry Pi 4B with a Streamlit interface integrating traffic capture, preprocessing, inference, visualization, resource monitoring, and automated alerts. The results demonstrate the feasibility of combining complementary detection strategies in an operational Edge-based IDS, while highlighting the trade-off between detection performance and computational cost.

Future Internet doi: 10.3390/fi18100535

Authors: Rashed Alsakarnah Mohammad Masoud

Diabetes-prediction models based on machine learning (ML) are often affected by inconsistent feature importance rankings. Different feature-selection algorithms emphasize different data characteristics and produce divergent ranked feature sets. Relying on a single ranking method can therefore lead to biased, unstable feature selection. To address this limitation, this work proposes Diabetes Hybrid Greedy Forward Selection (D-HGFS), a hybrid ML three-layer diabetes-prediction framework that integrates multi-perspective feature ranking with robust imbalance handling and classification. In the first layer, a hybrid model of ADASYN and the Tomek-links approach is utilized to tackle the issue of imbalanced classes. Subsequently, the second layer consists of seven heterogeneous feature-ranking algorithms from four categories: domain knowledge, filter, wrapper, and XAI-based methods. These methods are applied in parallel for fast execution. The D-HGFS algorithm aggregates these rankings to select the most informative features. This algorithm attempts to reduce the number of utilized features in one hand and enhance the prediction accuracy in the other hand. D-HGFS employs seven heterogeneous feature-ranking methods followed by a nested greedy feature-selection procedure. A leakage-free repeated nested cross-validation framework with five outer folds repeated 10 times was used, resulting in 50 outer-test evaluations. All preprocessing, imbalance handling, classifier selection, feature ranking, and feature selection were restricted to the corresponding training partitions. The final performance of the proposed framework was then exclusively evaluated on untouched outer-test folds. Finally, the third layer consists of a binary XGBoost classifier. Experimental results show that the proposed framework outperforms baseline XGBoost models while reducing the feature set by 42%. These findings demonstrate that combining multiple feature-ranking perspectives with effective imbalance handling produces a robust, accurate, and clinically efficient diabetes-prediction model.

Future Internet doi: 10.3390/fi18100534

Authors: Cossi Blaise Avoussoukpo Amara Camara Gervais Mwitende

Opportunistic networks acquire external communication, computation, sensing, storage, and service resources as mission requirements evolve. Prior research has developed human-mobility models, context-aware routing, trust mechanisms, and adaptive resource-management approaches, but these contributions address narrower decision scopes and do not by themselves provide an ergonomics-guided process for governing heterogeneous Helpers throughout mission execution. This paper formalises Ergonomic Opportunistic Networks (ErOppNets) as a technology-independent extension of mission-driven opportunistic networks in which ergonomics serves as a design compass for selecting and adapting technologies, policies, resources, and participation mechanisms to support human factors and ergonomics goals alongside mission performance. ErOppNets integrate human-centric, context-aware, trust-aware, and resource-aware decisions across discovery, assessment, admission, role allocation, monitoring, reassessment, reconfiguration, replacement, and release. A multidimensional orchestration model formalises Helper selection, while the ErOppNet Adaptive Orchestration (EAO) procedure translates the model into operational decisions. The evaluation compares six policies across nominal, dynamic, adversarial, resource scarcity, failure, and scalability scenarios using 100 paired random seeds per primary comparative configuration, complemented by separate parameter-sensitivity analyses. Under nominal conditions, EAO and Static Multidimensional (SM) selection achieved 100 per cent mission completion; relative to SM, EAO reduced mean configuration churn by 60.0 per cent. Across dynamic conditions and external populations up to 2000 entities, EAO preserved mission feasibility while balancing utility, stability, trust requirements, and adversarial exposure. The results support ErOppNets as a human-centric, adaptive, and technology-independent model for intelligent opportunistic network management.

Future Internet doi: 10.3390/fi18100533

Authors: Molka Gharbaoui Gabriele Cecchetti Sujit Kumar Sahu Anna Lina Ruscelli Piero Castoldi

Remote patient supervision and assistance play a vital role in modern healthcare, particularly in emergency scenarios where specialists may not be immediately available. To address this need, tele-assistance systems capable of delivering real-time medical guidance have become increasingly important, with emerging network technologies like 5G offering promising enhancements in terms of connectivity, latency, and data throughput. In this study, we deploy and evaluate a 5G-enabled tele-assistance solution designed to improve medical emergency response. The system consists of wearable helmets equipped with live-streaming capabilities, allowing on-site healthcare operators to transmit video feeds directly to remote doctors stationed at Emergency Service Operation Centers. To assess the benefits of 5G connectivity, we upgraded the existing system, previously reliant on 4G, to 5G and conducted an experimental campaign using the 5G Test Network (5GTN) at the University of Oulu. In addition, the helmet was upgraded with a Raspberry Pi 5 platform to improve the overall system performance. The experimental evaluation showed that the network and hardware upgrades increased the average throughput from 27.69 Mb/s with 4G to 129.9 Mb/s with 5G SA and reduced the round-trip time, providing a larger margin for real-time video streaming in emergency scenarios.

Future Internet doi: 10.3390/fi18100532

Authors: Md Monjurul Karim Qiang Qu

Autonomous AI agents discover services, negotiate, and settle payments. However, open agent markets lack reliable means to verify delivery, constrain delegated authority, and make identity renewal economically consequential. To address these failures, we formalize the verifiable agent economy, where evidence and programmable constraints govern identity, behavior, and payment. We introduce the agentic economy stack and formulate trust-asymmetric coordination through four requirements: delivery incentive compatibility, verifiable settlement, budget and authority safety, and Sybil resistance. To operationalize these requirements, we construct a blockchain-based mechanism combining verifiable credentials, stake, reputation, constrained wallets, and receipt-conditioned settlement. Within the class of receipt-conditioned payment rules, its geometric-mean payment rule minimizes worst-case multiplicative exposure to residual value uncertainty. The mechanism satisfies delivery incentive compatibility and Sybil resistance under explicit deterrence and identity-cost conditions, while its budget and authority guarantee is limited to actions routed through the enforcement path. Across five detection laws, numerical analysis identifies an interior verification-budget optimum. At their respective optima, our mechanism achieves 46% higher welfare than Verification-Only and raises honest delivery from 77% to 99%. A measured sampling-audit verifier shifts but preserves the optimum, while false positives and thin markets tighten participation and staking requirements.

Future Internet doi: 10.3390/fi18100531

Authors: Abdulmohsen Mutairi

Next-generation wireless communication systems have emerged as a significant means of supporting heterogeneous Internet of Things (IoT) applications that have different Quality of Service (QoS) requirements. Some issues, such as dynamic resource allocation, heterogeneous traffic handling, and adaptive orchestration, have been observed in existing approaches to network slicing. This work aims to develop an intelligent next-generation network slicing optimization framework capable of enhancing resource utilization, QoS provisioning, and adaptive slice orchestration through the integration of stochastic traffic modeling. The proposed framework uses a simulated 5G/6G network slicing dataset with heterogeneous IoT traffic. Initially, massive machine-type communication (mMTC), ultra-reliable low-latency communications (URLLC), and enhanced mobile broadband (eMBB) traffic are generated and monitored to collect network states. The collected data are normalized and processed using a deep autoencoder to extract meaningful features. These features are then used by Multi-Agent Proximal Policy Optimization (MAPPO) to optimize bandwidth, CPU, and memory allocation across slices, while Software-Defined Networking (SDN) and Network Function Virtualization (NFV) enable dynamic slice orchestration and adaptive resource management. Experimental results demonstrate that the proposed framework improves throughput by 20.9%, reduces latency by 49.1%, improves reliability by 3.4%, reduces packet loss by 69.2%, increases resource utilization by 9.6%, and improves SLA compliance by 19.4% compared with the baseline static resource allocation approach. Consequently, the proposed framework offers a robust and adaptive network slicing solution for next-generation IoT ecosystems.

Future Internet doi: 10.3390/fi18100530

Authors: Iacovos Ioannou Susanne Naegele-Jackson Sonja Filiposka Vincent Burkard Michael Georgiades Michael Mackay Vasos Vassiliou

Through intent-based networking (IBN), network objectives are specified declaratively, yet in mobile networks their realisation spans administrative domains that compete for coupled radio, transport and compute resources but cannot share local models and telemetry. A distributed artificial intelligence (DAI) controller with a Belief-Desire-Intention-eXecution (BDIX) cycle is presented, centred on shared-space coordination. The local decision of each domain agent is mapped through Gi into a common p-dimensional resource space, so that only resource contributions and a common price vector cross domains, after a minimum-slack feasibility certificate and behind a trust-admission gate. Typed refinement, deficit feedback and conflict mediation support it. In a coordination-layer benchmark (4 to 64 domains, twenty seeds), the full-information objective is matched to the reported numerical precision without disclosure of local models (cost gap 0.00%), whereas, at the principal β=0.60 setting, aggregate-only allocation loses 2.27% to 3.83% and equal split loses 25.21% to 37.67%. When 12.50% to 25.00% of the domains understate their contributions fivefold, honest-domain cost rises by 24.48% to 50.95% without the gate and by at most 2.26% with it, while mild falsification is not reliably detected by the trust gate, leaving containment dependent on the delegated resource bounds. On the single-domain analytical plant (twenty hold-out seeds), the controller attains the joint-highest displayed score (0.80), the lowest mean latency and significantly higher scores than seven of eleven comparators; end-to-end workflow execution is demonstrated separately by bridge dry runs. Shared-space coordination is thus the allocation of choice where local models cannot be shared.

Future Internet doi: 10.3390/fi18100529

Authors: Tajdid Ul Alam Mohamed Younis

Drone swarms operate collaboratively and serve critical missions, which require a common session encryption key to secure exchanged data. This key must be updated when new drones join—or existing ones depart—the swarm. Classical key management protocols force a trade-off between multi-message re-key chains and high broadcast cost, and they often degrade under packet loss and jamming. This paper presents PUG, a physically unclonable function (PUF)-anchored unified group key management protocol for the Internet of Drones that completes every membership event—join, leave, eviction, or re-key—with a single self-contained authenticated broadcast. The group controller encrypts the new session key separately for each drone under secrets derived from their PUFs. The controller then treats each node-specific ciphertext as a Chinese Remainder Theorem residue under a prime modulus that is node- and time-specific and derived from the PUF response. Solving this system of congruences yields a group lock, which is broadcast to all members in one message. Long-term private key material is never written to the drone’s nonvolatile memory, and a captured node is evicted from the group to keep future communication secured within the group and to prevent durable secrets that could be used to recover past group traffic. Membership update ensures forward and backward secrecy at epoch granularity. The simulation results show that our protocol is effective across various loss rates, maintains the largest share of the swarm keyed under sustained jamming, and imposes limited control traffic while being efficient in terms of communication and computation overheads.

Future Internet doi: 10.3390/fi18100528

Authors: Haris Alibašić

Agentic artificial intelligence (AI) shifts digital government from systems that generate recommendations toward networked systems that perceive, plan, communicate, invoke tools, initiate actions, and adapt with limited direct supervision. This article examines agentic AI as an intelligent cybersecurity governance problem in public administration because autonomous agents may access sensitive systems, execute delegated authority, propagate errors across Internet-enabled infrastructures, and create accountability gaps at machine speed. It uses a structured integrative review of 40 scholarly and technical sources and 15 formal governance instruments, together with a design-oriented synthesis and comparative framework analysis. The analysis identifies seven risk mechanisms: authorization drift, cross-system cascade, attribution fragmentation, observability loss, velocity-reversibility mismatch, adaptive-change risk, and contestability erosion. In response, the article develops the Hybrid Intelligence Cybersecurity Governance Architecture (HICGA), which links legal mandate and institutional responsibility to delegated action envelopes, consequence-based approval gates, zero-trust identity and access controls, action-chain logging, continuous assurance, interruption, rollback, and rights-preserving review. HICGA reframes nominal human-in-the-loop supervision as retained human authority. The architecture is a conceptual design proposition that still requires expert, simulation, and field validation. Autonomous administrative action is legitimate only when technological agency remains bounded by public authority, auditable across organizational and network boundaries, reversible where feasible, and open to explanation, contestation, and remedy.